Paul A. Sandery

Adaptive nonlinear dynamical initialization

[1] An adaptive nonlinear initialization scheme is presented that provides improvements over linear relaxation methods for approaching the target state while minimizing discontinuities in dynamical models. The method adds an adaptive nonlinear forcing term to the model equations that is a function of the difference between the model field and its target value. A new feature is that the amplitude of the forcing is nonlinearly adapted to the size of the difference, allowing for stronger relaxation where differences are large and weaker relaxation where differences are small. We find that the function leads to more optimal introduction of new information by working hardest at the beginning of the initialization period while converging toward a steady condition for the majority of the domain at the end of the initialization period. Experiments with a limited area ocean model, with different dynamical regimes, show that the adaptive scheme leads to less shock than standard linear approaches and permits the model to converge to a state away from the target field if the target is not a priori dynamically balanced. Results indicate that the method has the potential to lower forecast error. This suggests that it will have a broad range of applications in dynamical prediction systems.

Ocean forecasting of mesoscale features can deteriorate by increasing model resolution towards the submesoscale

Submesoscale dynamics are ubiquitous in the ocean and important in the variability of physical, biological and chemical processes. Submesoscale resolving ocean models have been shown to improve representation of observed variability. We show through data assimilation experiments that a higher-resolution submesoscale permitting system does not match the skill of a lower resolution eddy resolving system in forecasting the mesoscale circulation. Predictability of the submesoscale is inherently lower and there is an inverse cascade in the kinetic energy spectrum that lowers the predictability of the mesoscale. A benefit of the higher-resolution system is the ability to include information content from observations to produce an analysis that can at times compare more favourably with remotely sensed satellite imagery. The implication of this work is that in practice, higher-resolution systems will provide analyses with enhanced spatial detail but will be less skilful at predicting the evolution of the mesoscale features.The degree to which increasing the resolution of ocean models to consider submesoscale dynamics will improve prediction of mesoscale features remains uncertain. Here, via data assimilation experiments, the authors show higher resolution models do not necessarily provide improved dynamical solutions.

An adaptive quality control procedure for data assimilation

We describe a simple adaptive quality control procedure that limits the impact of individual observations likely to be inconsistent with the state of the data assimilation system. It smoothly increases the observation error variance depending on the projected increment, state error variance and so-called K-factor so that the resulting increment does not exceed the estimated state error times K. Because an estimate of the state error is readily available in the Kalman filter (KF), the method is particularly suitable for the KF, ensemble Kalman filter (EnKF), or ensemble optimal interpolation systems. The tests show that setting K to about 1.5–2 or above has no detrimental effect for performance of nearly optimal systems; at the same time it still makes it possible to make use of observations that might otherwise be discarded by the background check. The technique is successfully used in the EnKF codes TOPAZ and EnKF-C.