The rise of machine learning in climate modelling

Climate models are numerical representations of the Earth's climate system, including components such as the atmosphere, ocean and land. By solving mathematical equations based on physical principles, climate models explore long-term changes to the underlying statistical distributions that govern day-to-day weather (e.g. Mitchell and Scaife, 2020). Climate models are essential tools in understanding how human activity affects the climate, underpinning climate policy at all scales, from local to international. For informed decision-making, it is imperative that climate models are accurate and trustworthy. Development has typically come via improved scientific understanding and increased computational power, resulting in the current state-of-the-art climate models, Earth System Models (ESMs). Run on large supercomputers, ESMs can represent an increasing complexity of interconnected Earth system components at gridded global resolutions of 50–100km, and even finer resolutions approaching 1km when applied regionally. However, with the rate of conventional computational power slowing as the physical limits of transistors are reached, ESM development which relies on finer resolutions and handling large and/or complex datasets risks plateauing (Balaji, 2021). To continue advancing in these areas, such as improving the representation of small-scale processes and better capturing local impacts and extreme events, the modelling community will need to adopt alternative methods. One possible approach is to leverage machine learning (ML). A subset of 'artificial intelligence', ML uses computer algorithms that discover patterns and trends within training data to form models capable of prediction from 'unseen' data. Many ML algorithms use established statistical methods, such as linear regression, clustering and optimisation; generally, it is the purpose behind using a method that would classify it as ML or not. ML often prioritises generating accurate predictions of a system over making inferences of the underlying properties – in statistics, the reverse is usually true. The ability of ML to scale and adapt to large and/or complex datasets, as well as harness the efficiency of graphical processing units (GPUs), has led to its increasing prominence within weather and climate modelling (de Burgh-Day and Leeuwenburg, 2023). Although current successes are largely associated with weather forecasting, the crossover between the two disciplines suggests that similar progress is possible in climate modelling. ML could provide a step-change in climate modelling. For exampe, with ML, global models representing Earth's processes at resolutions finer than 1km with great accuracy and speed are a real possibility. Such capabilities could realise long-standing ambitions such as real-time climate projections, discovery of unknown Earth system processes, enhanced regional/local impact and adaptation assessments, operational extreme event attribution and improved accessibility to climate data (e.g. de Burgh-Day and Leeuwenburg, 2023; Eyring et al., 2024). To achieve such feats will require significant expertise and infrastructure. Whereas academia can offer the former, the private sector currently has the latter, thus future collaboration between both will be crucial if ML is to advance climate modelling. Although the use of ML in climate models is still in its infancy, a new era is approaching – and it may arrive sooner than you expect. This work and its contributor (George Jordan) were funded by the Met Office AI4 Climate project which is supported by the Department for Science, Innovation & Technology (DSIT).