• Evaluation and uncertainty estimation from multimodel and perturbed physics ensembles (Reto Knutti, ETHZ)

    • Climate change projections and uncertainties are often based on a <91>one model, one vote<92> approach to determine an average and spread of all models, despite the fact that they differ in terms of resolution, number and complexity of included processes, forcings, and agreement with observations. The models considered are part of multimodel ensembles, or perturbed physics ensembles, i.e. versions of a single model run with multiple sets of parameters. By construction, the spread of multimodel ensembles of opportunity does not reflect the full range of current model uncertainty and its average is biased towards those groups that submit more data and model versions. Given that some but not all of the newer models now include for example interactive representations of the carbon cycle, nitrogen cycle, ice sheets or land use, a simple model average and spread is increasingly difficult to defend and to interpret. With the CMIP5 model intercomparisons around the corner, the climate modeling community is struggling to answer the following questions, among others:

  • What is a good model? What metrics are most useful for model evaluation? Are metrics evaluating the ability of a model to simulate the current mean climate, its variability, or trends useful to constrain the future prediction?
  • Is it useful to define an overall metric of skill for a model, or should each application have a separate metric?
  • Is weighting of models justified, and how?
  • In which applications have methods to weight or rank models been useful or successful (or failed) and why?
  • What methods can be used to combine perturbed physics ensembles and structurally different models?
  • Are multimodel and perturbed physics ensembles broad enough to start with? Is structural error (the fact that many models are wrong in a similar way) a problem?
  • Is uncertainty best expressed in terms of probabilities over all the models, or as a few illustrative cases selected from a set of models?
  • At which temporal and spatial scale can the models provide useful information? At what point should we stop quantifying uncertainty and providing predictions and simply tell the user that we do not know? When will we know more?
  • What part of the uncertainty can be reduced, which part is irreducible?

    This session welcomes contributions addressing the above topics from statistical but also climate process point of view. Contributions from other areas (e.g. seasonal forecast) where model evaluation and uncertainty is critical, as well as and contributions discussing user needs are also welcome.

    Invited Speakers

  • Andreas P. Weigel - Multi-model combination on seasonal and multi-decadal time-scales
  • Alex Hall - A strategy to improve projections of Arctic climate change
  • David Sexton, Ben Booth, Mat Collins, Glen Harris and James Murphy - Using a perturbed physics ensemble to make probabilistic climate projections for the UK
  • Ben Sanderson - Perturbed physics and multimodel ensembles: How can we use them together to constrain future climate response?