Learning from the past to improve ocean decadal climate predictions through hindcasts

Learning from the past to improve ocean decadal climate predictions through hindcasts

SEACLIM is developing regional ocean predictions to help improve understanding of how Europe’s seas may change over the coming years and decades. Before these predictions can support future climate information, they need to be carefully tested.

This is where hindcasts become essential. A hindcast is a prediction made for a period in the past. By comparing what numerical models predicted with what was measured by scientific instruments, researchers can assess whether a given model simulation can reproduce known ocean variability and events.

SEACLIM uses hindcasts to test and improve regional ocean multi-year prediction systems. This work helps assess prediction skill, refine downscaling approaches, and better understand how past and present ocean conditions influence future predictability. 
This news article focuses on SEACLIM’s work on using hindcasts to test and improve multi-year ocean predictions. Short-term operational forecasts and long-term climate projections are not addressed here.

 

The hindcast concept. Hindcasts are predictions started in the past and compared with historical observations or reference data to assess model skill.

 

Why hindcasts matter

Could ocean temperatures be unusually high over the coming years? Could changing marine conditions affect ecosystems, aquaculture, or coastal risk? Ocean predictions can help scientists and decision-makers plan, but they always come with uncertainty.

The key question is therefore not whether a prediction is always right, but how reliable it is. One way to assess reliability is to produce a prediction, wait until the predicted period has passed, and compare it with what happened. A more powerful approach is to start the prediction in the past and compare it with historical observations or reference datasets. This is a hindcast.

If a model can reproduce past variability and events in hindcasts, researchers gain confidence in its ability to provide useful information for future predictions. In SEACLIM, hindcasts play a vital role in testing and improving regional ocean prediction systems. They help assess prediction skill, refine downscaling approaches, and explore how ocean memory influences predictability.

To make these hindcasts reliable at the regional level, SEACLIM also needs a consistent way to downscale information from global climate prediction systems into regional ocean models. This is one of the core tasks of SEACLIM.

 

A framework for regional predictions in SEACLIM

What are the best practices for producing regional multi-year ocean hindcasts? This is still an active research question. Shortly after the SEACLIM kick-off, we gathered experts in multi-year ocean prediction to discuss how hindcasts should be downscaled from global models to high-resolution regional models.

A key part of this work is defining how regional models receive the information they need to run. Regional ocean models do not operate in isolation. They need information from outside the modelled area, such as atmospheric conditions above the ocean and ocean conditions at the boundaries of the region. Scientists call this information forcing.

Agreeing how this forcing is prepared and used is important because it affects the quality and comparability of regional hindcasts. One early SEACLIM outcome is therefore not only a set of simulations, but also a shared framework for running regional prediction experiments in a consistent way across the various regions covered by the project.

This framework supports future efforts to make the forcing needed for regional downscaling more routinely available from global prediction systems. Based on this framework, SEACLIM is now using established guidelines to run concrete hindcast experiments.

 

The four main European marine basins focused on SEACLIM. The work conducted in these regions is designed from the early days of the project to be replicated in other regions.

 

How SEACLIM performs hindcasts in practice

SEACLIM tests its prediction systems by repeatedly going back to the past. Our hindcasts are designed as a series of five-year predictions starting each November from 1993 to 2024. These simulations combine information from a global climate prediction model with regional ocean models covering the Arctic Ocean and subarctic seas (SEACLIM domain 2) and the North Sea, northeast Atlantic and western Mediterranean Sea (SEACLIM domain 1).

For each start year, several slightly different hindcasts are run. These minor differences in the initial conditions lead to different outcomes over time, creating an ensemble of evolutions. This ensemble approach helps represent uncertainty linked to internal variability in the climate system.

In practice, this means researchers do not rely on a single model run. By comparing several plausible evolutions of the ocean, they can better understand where predictions are more robust and where uncertainty remains higher. In SEACLIM, regional models downscale ensemble information from global climate prediction systems, such as the Norwegian Climate Prediction Model (NorCPM), which provide the initial conditions and forcing used to run the regional hindcasts.

The hindcasts are then compared with ocean observations and reference datasets to evaluate prediction skill. In simple terms, prediction skill describes how well the model reproduces past variations in the ocean. Statistical measures, such as anomaly correlation and root mean squared error, help quantify this skill.

As with weather forecasts, prediction skill usually decreases as the lead time increases. A prediction is expected to be more reliable close to its start date than several months or years later. Evaluating this performance is a crucial step before using the predictions for scientific analyses or decision-relevant information.

However, multi-year ocean predictions are different from short-term weather forecasts. They are not usually used to predict the exact value of an ocean variable on a specific future day. Instead, they help assess whether conditions over a period or region may be higher, lower, or different from what is typical.

 

SEACLIM regional domains used for numerical modelling downscaling activities.

 

How well do the predictions perform?

During the first 18 months of the 48-month project, SEACLIM has been evaluating state-of-the-art decadal climate prediction systems, including the Norwegian Climate Prediction Model (NorCPM), to understand how well they can predict ocean climate conditions relevant to SEACLIM. These analyses build on simulations contributed to the Decadal Climate Prediction Project (DCPP) within the World Climate Research Programme (WCRP) Coupled Model Intercomparison Project (CMIP6). Readers who would like to go deeper into the scientific framework behind these experiments can explore the overview paper, The Decadal Climate Prediction Project (DCPP) contribution to CMIP6.

Early analyses show that these systems can predict variations in ocean climate conditions, especially during the first year after initialisation. Surface air temperature shows good predictability in several regions, while sea surface temperature and salinity show strong predictability in the North Atlantic and in regions covered by SEACLIM activities.

Arctic sea ice also shows high prediction skill in some areas, particularly in the Chukchi and Beaufort Seas. Winter sea ice tends to be more predictable than summer sea ice. As expected, prediction skill decreases after the first year but remains positive for North Atlantic sea surface temperature in the systems assessed.

These results also show why ensembles and model comparison matter. Combining predictions from multiple models can improve prediction skill by reducing model-specific errors. Correcting for model drift, which is the tendency of a model to move away from observed climate conditions over time, can also improve prediction quality. Initialised predictions provide benefits compared with simulations that are not initialised from observed climate conditions.

Early analyses show that these systems can predict variations in ocean climate conditions, especially during the first year after initialisation. Surface air temperature shows good predictability in several regions, while sea surface temperature and salinity show strong predictability in the North Atlantic and in regions covered by SEACLIM activities.

Arctic sea ice also shows high prediction skill in some areas, particularly in the Chukchi and Beaufort Seas. Winter sea ice tends to be more predictable than summer sea ice. As expected, prediction skill decreases after the first year but remains positive for North Atlantic sea surface temperature in the systems assessed.

These results also show why ensembles and model comparison matter. Combining predictions from multiple models can improve prediction skill by reducing model-specific errors. Correcting for model drift, which is the tendency of a model to move away from observed climate conditions over time, can also improve prediction quality. Initialised predictions provide benefits compared with simulations that are not initialised from observed climate conditions.

 

Testing ocean predictions by looking back. The figure shows how hindcasts are used to compare modelled past conditions with observations and reference datasets. The examples shown use lead-year-1 initialised hindcasts from the CMIP6 DCPP model IPSL-CM6A-LR. The maps show the anomaly correlation coefficient (ACC) of drift-corrected sea surface salinity and near-surface air temperature over the North Atlantic. Red indicates higher agreement with reference data, blue lower agreement, and hatching marks results that are statistically significant at the 95% confidence level. The time series compares drift-corrected sea surface temperature in SEACLIM domain 1 with HadISST observations, CMIP6 historical simulations, the hindcast ensemble mean, and individual ensemble members, helping researchers assess the added value of initialisation.

 

Predicting biogeochemical conditions

The physical ocean, including temperature, salinity, and currents, strongly influences biogeochemical processes such as air-sea CO2 fluxes, pH, and marine primary productivity. Because of this connection, predictability in the physical ocean can translate into some degree of predictability in ocean biogeochemistry.

In SEACLIM, early results indicate that the global prediction model can capture changes in ocean nutrient concentrations several years ahead. Primary productivity, however, is harder to predict. This highlights both the limits of current systems and the areas where regional modelling may help improve future predictions.

Regional ocean models can represent smaller-scale processes that are not captured in the same way by global models. This may help bridge some of the gaps in predicting biogeochemical conditions, especially in coastal and shelf-sea environments where local processes can strongly influence ecosystems.

 

Plankton bloom off the coast of Norway. SEACLIM aims to improve our ability to assess the biogeochemical conditions that can contribute to events such as harmful algal blooms, supporting the regional-to-local downscaled information that aquaculture operators and marine protected area managers need on the ground. Credits: European Space Agency.

 

Putting predictions to use

SEACLIM's research on regional marine predictions will feed into later project activities, including the development of essential ocean variables, ocean climate indicators, and climate service demonstrators. Essential Ocean Variables are standard variables used to describe key features of the marine environment, such as temperature, salinity, oxygen, or biological productivity.

For each variable, SEACLIM will be able to evaluate how far in advance it can be predicted and compare the prediction skill of global and regional models. This will help identify where regional downscaling adds value and where further improvements are still needed.

Stakeholders connected to SEACLIM include coastal authorities, port actors, aquaculture companies, and other marine users. Multi-year predictions could help assess the likelihood of events such as marine heatwaves or extreme sea-level conditions. However, these predictions are currently research-level outputs. If shown to be robust, they could contribute valuable information on coastal risks and marine conditions for sectors operating in the marine environment.

Hindcasts remind us that predicting the future depends on understanding and accurately reproducing the past. By systematically assessing how well models reproduce past ocean variability, SEACLIM is helping make multi-year ocean predictions more robust and more useful for those who rely on regional ocean climate information.

 

SEACLIM’s main stakeholders and end users come from various key international climate and ocean data programmes and frameworks.

 

Continue exploring SEACLIM

To learn more about SEACLIM and its work on regional ocean climate prediction, visit the project website and read the various stories that the project already published. Follow us on LinkedIn to ensure that you do not miss any important announcement, new story, or opportunities to engage with SEACLIM.

 

The SEACLIM project is funded by the European Union under the Horizon Europe Research and Innovation programme of the European Research Executive Agency (REA). Any posts reflect only the views of the project.