"We need to reconsider the design of climate model experiments and greatly increase the sampling of initial condition uncertainty," Joseph Daron of the University of Cape Town and David Stainforth of the London School of Economics, University of Warwick and University of Oxford told environmentalresearchweb. "This is the case whether our aim is to understand our models, support scientific understanding of the climate system, or provide information that is valuable in guiding policy decisions."

Daron and Stainforth analysed the performance of climate models by using a much simpler version, an idealized low-dimensional nonlinear chaotic model. The model was analogous to a coupled ocean-atmosphere system as it contained both a slow-varying component, to mimic the ocean, and a fast-varying atmosphere-like component. But it needed much less computer processing power than a coupled ocean-atmosphere climate model.

"Using such models we are able to run multi-thousand member ensembles to explore the impact of varying initial conditions and model parameters on the evolution of the system," said Daron. "Results from such experiments cannot be assumed to provide quantitative climate prediction information, but they can provide qualitatively relevant output that helps us understand the limitations of our assumptions and of current modelling approaches."

The team found that single simulations or small ensembles of simulations of their simple model could not provide reliable quantifications of the model’s climate distributions.

"The result is found both in stationary forcing conditions and non-stationary forcing conditions," said Stainforth. "This implies that for seasonal to decadal predictions we ought to run large initial condition ensemble experiments, irrespective of whether the climate is changing; under climate change, assumptions about the ability of single simulations to quantify climate distribution simply get worse."

Traditionally, researchers have assumed that the uncertainty in initial conditions is relatively unimportant over timescales of decades. "The initial condition ensembles in CMIP3 [Coupled Model Intercomparison Project 3] and CMIP5 are likely too small to reliably quantify their climate distributions and are therefore unable to characterize future climate, even within their own ‘model-worlds’," writes the team in Environmental Research Letters (ERL). CMIP5 specifies a minimum of three initial condition ensembles.

If this source of uncertainty is not well quantified, the researchers continue, multi-model-based probabilistic climate predictions will inherit the errors in the distributions of each model. As a result, their probabilities may not be reliable for use in planning. Without awareness of errors from initial conditions, for example, developers of a river flood protection scheme using probabilistic predictions for specific rainfall thresholds could either over-adapt, leading to extra expense, or under-adapt, giving increased risk.

"We frame our work in the context of real-world adaptation and mitigation decision problems," said Stainforth. "This means we therefore attempt to bridge the nonlinear dynamics, climate modelling and adaptation communities, making our findings relevant for those who are tasked with basing decisions on estimates of future climate risk using climate model output."

As well as arising from initial conditions, uncertainty in climate model projections can be due to uncertainty about future greenhouse gas concentrations (forcing uncertainty), or to man’s ability to represent the Earth system in a computer model – model uncertainty.

Given practical computational constraints it is expensive and time-consuming to run an operational climate model more than a handful of times, said Daron. But as computational capacity increases, the possibility of running large ensembles will become more feasible. "In order to ensure that such future climate model experiments are informative for both science and society, we need to understand how we can best design the experiments to explore the different sources of uncertainty and how to provide meaningful quantitative output to inform probabilistic climate risk assessments," he added.

Now the researchers are extending their results and investigating operational climate models. "In addition, research in the adaptation decision-making domain is continuing to understand further the information needs of different sectors in diverse contexts," said Daron.

Related links

Related stories