Artificial probabilistic information for ocean-climate applications – REPLICA

Ongoing climate change is a critical problem for society, with wide-ranging potential impacts. Through its ability to store, transport and release heat, the ocean will play a key role in determining future climate evolution by modulating the atmospheric warming response. However, the ocean is able to generate unpredictable, intrinsic variability at multiple temporal and spatial scales, which can complicate the detection and attribution of this long-term climate change in many observed quantities. The probabilistic information that is needed to resolve this problem can, at present, only be obtained by ensemble modelling, but this approach has both scientific and technical flaws, including the dependence of the results obtained on model physics and spatial resolution, and the immense computational resources that are required. This immense demand on computational resources limits the possible model configurations that can be used in practice, constraining both the ensemble size and the model’s spatial resolution. These constraints may affect the results that are obtained. A further disadvantage is that the associated production, storage and analysis of the ensemble have a substantial carbon footprint.

REPLICA will apply statistical and machine learning techniques to a variety of oceanic data sets with the aim of improving our understanding of the impact of intrinsic ocean variability on climate change metrics, and a focus on the North Atlantic. This region is critical in determining European climate, but is strongly influenced by intrinsic variability through the Gulf Stream / North Atlantic Current. We will seek to do achieve this improved understanding by: (i) enhancing existing probabilistic information, and (ii) developing accessible and sustainable methods of obtaining probabilistic information that do not depend on direct numerical simulation.

We will enhance existing probabilistic information by developing methods to quantify the effects both of model physics and of limited ensemble sizes on probability density distributions estimated using ensemble model experiments. We will develop statistical techniques to estimate the ensemble mean directly from observations, and to refine the probability density estimates in regions where the intrinsic variability is strong. By applying these methods, we will seek to attenuate the influence of intrinsic variability in observation-based estimates of ocean heat and freshwater content variability in the North Atlantic, and to better understand the mechanisms of ocean-atmosphere interactions in the Gulf Stream region.

In parallel, we will seek to develop methods of obtaining probabilistic information that do not depend on direct numerical simulation through the application of artificial intelligence techniques. We will apply these methods to high resolution data sets, with the aim of better understanding the sensitivity of the processes governing deep convection to atmospherically-driven and intrinsic variability. This process is important in determining the strength of the overturning circulation, but remains difficult to predict.

The tools that we propose to develop in REPLICA will provide a less computationally-intensive and technically-demanding alternative to explicit numerical ocean simulation, with reduced demands on data storage. This will improve the accessibility of probabilistic information, facilitate climate change detection/ attribution tasks, and reduce the energetic cost of both production and storage of this information. We anticipate that the associated data sets will permit new insights into climate-critical processes in the North Atlantic.