Estimation of the time-varying probability density function from ensemble simulations and observations using Analogs

Understanding the role of ocean-atmosphere interactions is crucial in determining the drivers of ocean variability. Indeed, a part of this variability is not driven by the atmosphere but spontaneously and randomly generated by the ocean through non-linear processes. This internal variability is associated with multiple spatial and temporal scales, and may complicate the detection and attribution of climate change signals. Hence, quantifying the relative importance of atmospherically-forced and chaotic intrinsic variability is necessary to understand the mechanisms of climate change in the ocean-atmosphere system. However, both atmospherically-forced and intrinsic variability cannot be estimated with a single model experiment alone : an ensemble simulation approach is required. The ensemble mean approximates the component of the oceanic variability that is due to the influence of the atmosphere, while the spread represents the range of the estimated intrinsic variability. This work investigates the possibility of describing and predicting the random part of the ocean's variability from observations using an ensemble of ocean simulations in the North Atlantic ocean. An analog-based method is developed, and applied to Sea Surface Height data, with the aim of obtaining a less-computationally expensive method of estimating the time-varying probability function (PDF) that is normally obtained through ensemble simulation. The ensemble is supplied by the multi-decadal (1960-2015) global ocean/sea-ice eddy-permitting (1/4° resolution) large (50-member) ensemble simulation (OCCIPUT Experiment). The ensemble of SSH data as a whole provides the target PDF that we seek to estimate in a regions representative of the diversity of flows in the North Atlantic (e.g. at the centre of the North Atlantic gyre and in the Gulfstream current). The individual members are used to form the catalog of simulations in order to find analogs situations on which the estimate of the target PDF is based at time t. First results are promising and show that we are able to estimate the ensemble mean, but the variance is still a subject of active work due to the complexity of the shape of the PDF. The method greatly reduces the time and resources of computation by producing mean and variance of time-varying PDF for the entire time series in generally a few tens of minutes.

keywords : Internal variability, detection and attribution, model uncertainty, ocean-atmosphere interaction, predictability