Skip to main content

Synchronisation to enhance reliability of climate predictions

Periodic Reporting for period 4 - STERCP (Synchronisation to enhance reliability of climate predictions)

Reporting period: 2020-03-01 to 2021-08-31

Climate prediction is the next frontier in climate research. Prediction of near-term climate is limited mainly to the ocean. Although there is confidence in predicting long-term changes in climate at global scales, large uncertainties exist in future climate change, particularly at regional scales. Thus, providing climate services for decision making and informing policy is a challenge, and improving climate prediction is a high priority, as society requires reliable short and long-term prediction of climate.

STERCP has investigated the potential of an innovative “supermodel” technique to reduce model systematic error, and hence to improve climate prediction skill and reduce uncertainties in future climate projections. The current practice to account for model systematic error, as for example adopted by the Intergovernmental Panel on Climate Change (IPCC), is to perform simulations with ensembles of different models. The problem with this approach is that common long-standing model errors are not reduced. Instead of running models independently, we propose to connect the different models in a manner that they synchronise and errors compensate, thus leading to a model superior to any of the individual models – a supermodel. This concept stems from theoretical nonlinear-dynamics and relies on advanced machine learning algorithms.

STERCP achieved its overall objectives to develop the first climate supermodel using state-of-the-art Earth System Models (Fig. 1), and to demonstrate the potential of supermodelling to reduce model errors and thereby enhance climate prediction and reduce uncertainties in climate projections. Beyond this, the project has improved understanding of systematic model error and its impact on predicting climate.
STERCP has developed a supermodel based on three state-of-the-art ESMs by connecting their ocean model components, and by training it using global observations of sea surface temperature. A version with two of these ESMs connected via their atmosphere and ocean models was also constructed. Simulations of present day and future climate were performed.

This key achievement required developing new software and machine learning algorithms to efficiently connect and train supermodels. To connect ocean components we applied advanced data assimilation methods to synchronize the models on a common ocean state every month. To connect atmospheric components we ingested a common atmospheric state to each atmosphere every six hours. The common state is an optimal weighted sum of the individual model components. Two training schemes have been developed to determine the optimal weights: synchronised based learning and cross-pollination in time. Both schemes are designed to give larger weight to those models that most reduce the supermodel error.

Computationally more efficient supermodels were developed to facilitate this work. Supermodel behaviour was studied in an interactive ensemble configuration, in which different atmospheric models are connected to a single ocean model and their interaction with the ocean is optimised to reduce model error. Training schemes were developed using a supermodel based on an intermediate complexity climate model. Here the three-dimensional states of the different atmospheric models were connected, while each atmospheric model was again coupled to the same ocean model.

Inspired by supermodelling concepts, we also developed two new approaches to study climate variability. In the first, we developed a regional coupled interactive ensemble. This is a useful tool to extract patterns of climate variability caused by two-way interaction between the ocean and the atmosphere in a specific region. In the second, we developed a partial coupled approach to study the impact of wind driven ocean variability on the atmosphere, while maintaining consistent thermodynamic interaction between ocean and atmosphere.

Extensive numerical experiments were performed with the different supermodels and derived modelling approaches, as was wide ranging analysis of available multi-model ensemble simulations. We demonstrated that supermodelling can substantially reduce long-standing model errors in rainfall (Fig. 2) and sea surface temperature patterns across the tropics, and that the climatically important E l Niño phenomenon could be better simulated. We also showed the potential to improve climate prediction and long-term climate change projections using idealised experiments with computationally cheaper models. We also delivered key new insights on the detrimental effects of model errors on short and long-term climate predictions.

Our findings were disseminated through numerous peer-reviewed publications and international conference and workshop presentations. This has generated substantial interest in supermodelling, and invitations to participate in research proposals. Thus, we are confident in the further development and exploitation of supermodelling.
Supermodelling represents a step change in climate modelling. Following the current practices, eliminating long-standing errors will take decades, and requires much more powerful computers than presently available. For example, most climate models used in the last three IPCC assessment reports simulate an erroneous band of rainfall south of the equator in the Pacific and Atlantic, introducing large-uncertainties in climate change projections.

STERCP has demonstrated that such errors can be largely eliminated by combining current models using presently available computers. In particular, we constructed the first ever supermodel based on state-of-the-art ESMs. These were connected via their oceanic states using data assimilation. We trained the supermodel using observations of sea surface temperature. The resulting supermodel was able to largely eliminate errors in rainfall patterns (Fig. 2) that have persisted across generations of climate models. This level of error reduction could not be achieved by averaging independently performed simulations, as traditionally done. The supermodel better simulated the major patterns of tropical climate variability. The supermodel is superior because as an interactive ensemble it can extract benefit from non-linear ocean-atmosphere interaction.

We have demonstrated the potential to extend supermodelling to global climate prediction using an intermediate complexity supermodel with synchronised atmospheric states. We used an idealized configuration in which the supermodel was trained to reproduce a “perfect” model—a version of the same model with different parameters. The supermodel’s climatology and its short-term weather forecasts and long-term climate change projections were closer to the perfect model, as compared to the imperfect models and the standard multi-model ensemble average.

STERCP also introduced other novel modelling approaches to study climate dynamics and impacts of model errors. One major achievement resulting from these was to show that reducing model errors could substantially enhance seasonal prediction in the tropical Atlantic, a region where state-of-the-art models have little skill.

STERCP made great progress in demonstrating the potential of supermodelling to revolutionize climate modelling. The operationalization of supermodelling now requires the development of new technology that can combine climate models seamlessly, and ideally across the internet.
Fig. 2: Supermodel improves the simulation of rainfall
Fig. 1: Supermodel Schematic