Can we solve the “spin-up” problem in Earth System Modelling?

The challenge

Earth System Models (ESMs) are central to understanding the climate system and predicting its future evolution. They underpin IPCC assessments and inform major international policy processes, including the Global Stocktake. 

However, these models are extremely computationally expensive to run. This creates a significant constraint on the timeliness and scope of the evidence available to policymakers. 

Before simulations can be used, ESMs must first be “spun up” to a stable pre-industrial state. This process requires simulating thousands of years of climate dynamics to allow slow-moving components, such as the deep ocean and terrestrial carbon cycle, to equilibrate. Even on the most powerful supercomputers, a single spin-up can take more than two years. 

This limitation has direct consequences for both science and policy. In practice, it is often only feasible to run a single spin-up. This restricts the ability to explore the substantial uncertainties inherent in these models and narrows the range of future climate projections available to decision-makers. 

As a result, policymakers must rely on a constrained evidence base when making decisions on critical issues such as carbon budgets, climate risks, and long-term adaptation planning. Despite its significance, the spin-up problem remained unresolved for decades. 

Previous research by Khatiwala in 2023 demonstrated a promising numerical algorithm that was more than ten times faster than conventional approaches, but further work was needed to test and demonstrate its robustness in operational settings. The scale of the scientific and policy challenge, and the potential to unlock more reliable projections, created a clear demand for a practical solution, bringing the Met Office into the Sprint as a key partner to assess how the approach could be applied within UKESM and inform future IPCC-relevant modelling.

The Solutions

The Sprint delivered a scalable and operational solution to the long-standing spin-up problem by developing, testing and implementing a fast numerical algorithm across multiple components of Earth System Models.  

The team successfully: 

•  interfaced the algorithm with MEDUSA and JULES, the marine biogeochemical and land vegetation components of UKESM 

• adapted the code to run within the Met Office’s workflow software and supercomputing environment 

• demonstrated substantial acceleration in spin-up times, achieving performance improvements of approximately 10 to 12 times faster for NEMO-MEDUSA and between 10 and 50 times faster for JULES.  

• generated multiple equilibrium realisations of MEDUSA using parameter sets identified by the Met Office as key sources of bias in previous IPCC simulations, creating a pathway for improved representation of uncertainty in future projections 

They not only improved computational efficiency but also revealed important differences in model complexity, particularly highlighting the greater challenges associated with land system models. 

The pathway

The initial focus was on accelerating the spin-up of MEDUSA and JULES, the marine biogeochemical and land vegetation components of UKESM, which represent key rate-limiting processes in Earth System Models. Once this acceleration was successfully demonstrated, the algorithm and associated code were made available to the Met Office and adapted for deployment within their supercomputing environment for production-scale simulations. This transition from methodological innovation to operational application use was critical to ensuring that the approach could directly inform CMIP simulations and, by extension, future IPCC assessments. 

Together, these advances provided the Met Office and the wider international modelling community with a practical, transferable solution that could increase the range of scenarios explored in CMIP simulations and strengthen the evidence base underpinning climate policy decisions. 

A peer reviewed publication in Science Advances and an accompanying article in The Conversation established the credibility and visibility of the approach within the scientific and policy relevant modelling community.

What happened next?

The algorithm was consolidated into a transferable, open-source tool and successfully deployed on the Met Office supercomputer for production calculations. 

It was successfully tested beyond UKESM components, showing that the method generalises across a wide range of leading Earth System Models and components. This included the National Center for Atmospheric Research Community Land Model in the United States, the Norwegian Earth System Model, the PISCES biogeochemical model used within NEMO by multiple European and international modelling centres, and the Model of Oceanic Pelagic Stoichiometry at GEOMAR in Germany.  

The code was made openly available via Zenodo, enabling uptake and adaptation by the international modelling community, with early take up reported across modelling centres in Spain, the USA, Norway, Japan and Germany, reinforcing its potential as a shared methodological advance for reducing the time and cost of producing the climate projections policymakers rely on. An andacc user guide was developed to help practitioners implement the algorithm in their own workflows. 

  “Policymakers rely on climate projections to inform negotiations as the world tries to meet the Paris Agreement. This work is a step towards reducing the time it takes to produce those critical climate projections.” (Professor Helene Hewitt OBE, Co chair of the CMIP Panel)