HARMONY

As climate change continues, high quality climate information over the short to medium term (years to decades) is required in order to foresee, mitigate, and adapt to its impacts. Currently, much of this information is created using climate models of high complexity. As such, there are two key challenges in climate science: 1) Creating climate models that are of high quality; for example, balancing the improvements from increased complexity with increased computational expense and the exploration of potential sources of parameter uncertainty, and 2) Effectively exploiting the outputs of multiple models; for example, the projected 18 PB of data representing approximately 100’000 total simulated years on new multi-model archives (“Big Data”). In HARMONY, we propose a combined strategy to harness this data to both efficiently improve climate models and make better predictions and projections of future climate change over Europe. We are now at an inflection point, where these efficiencies and improvements can be made a reality. A combination of state-of-the-art new approaches in Machine Learning (in particular in interpretability) and Uncertainty Quantification are themselves made possible by new infrastructure: the vast, new 6th Coupled Model Inter-comparison Project (CMIP6) archive and a game-changing new GPU cluster at the Host Institution dedicated to Machine Learning. In HARMONY, we will integrate these new approaches and infrastructure to efficiently find and exploit the high quality - but often hard to access - information within climate models. The primary outcomes of HARMONY will be threefold: 1) Development of new Machine Learning approaches and Interpretability tools, resulting in greater understanding of European climate variability, leading to 2) A demonstrable framework for the efficient improvement of climate models, yielding more robust climate projections, and 3) Increased skill in - and understanding of - near term predictions of European climate.