All models are wrong, but some are useful, as the saying goes. However, when looking at how Arctic sea ice decline is modeled, one might be tempted to say that all sayings are useful, but some are wrong. To be fair, I should be the last person taking a piss at climate models. Hundreds of brilliant scientists, engineers and IT specialists are giving their best every day to make supercomputers come up with scenarios that project future changes. Unfortunately, there is no Planet B to experiment on.
But we have come to a point where fake skeptics show up in television programmes (such as last week's BBC Newsnight) and use modeled predictions for Arctic sea ice in the 2007 IPCC Fourth Assessment Report as an argument not to be worried about the disappearance of Arctic sea ice, because "none of them shows it melting before the year 2070 on a regular basis in the summer" (quote from UK Conservative MP Peter Lilley).
As always, they're not telling the whole story:
This image (taken from the Climate Crocks blog) comes from a GRL research paper by Stroeve et al. that was published in 2007, detailing how models that participated in the World Climate Research Programme Coupled Model Intercomparison Project Phase 3 (CMIP3) were doing compared to observations (red line). The 2011 dot was drawn in, based on NSIDC data on September 7th of last year. Commenter Tim took the liberty of drawing in the new 2012 record. This picture saves us the 1000 words needed to explain how off models are when it comes to matching observations, which essentially makes their projections worthless.
But as the Arctic sea ice cover has changed radically in just 6 years, so have new and improved models started a new round of simulations and projections for the Phase 5 project (CMIP5) for the next IPCC assessment report. And again, Dr. Julienne Stroeve and her colleagues from the NSIDC, NCAR and the Voeikov Main Geophysical Observatory in Russia, have taken a look at how these models are faring in a new GRL paper that was published three weeks ago in Geophysical research Letters: Trends in Arctic sea ice extent from CMIP5, CMIP3 and observations.
From the introduction:
The rapid retreat and thinning of the Arctic sea ice cover over the past several decades is one of the most striking manifestations of global climate change. Previous research revealed that the observed downward trend in September ice extent exceeded simulated trends from most models participating in the World Climate Research Programme Coupled Model Intercomparison Project Phase 3 (CMIP3). We show here that as a group, simulated trends from the models contributing to CMIP5 are more consistent with observations over the satellite era (1979–2011). Trends from most ensemble members and models nevertheless remain smaller than the observed value. Pointing to strong impacts of internal climate variability, 16% of the ensemble member trends over the satellite era are statistically indistinguishable from zero. Results from the CMIP5 models do not appear to have appreciably reduced uncertainty as to when a seasonally ice-free Arctic Ocean will be realized.
This is a very interesting paper that is easy to follow. Two metrics are used to evaluate the performance of CMIP5 climate models: 1) The distribution of simulated extents over the period of observations (1953-2011) is used to assess how well the models capture the observed state of the ice cover, and then 2) Trends in simulated ice extent are used as a measure of the ability of the models to capture the response of the ice cover to global climate change. Quote:
It is possible that a model can capture the historical state but not the trend. We compare results of these evaluations to those based on the CMIP3 simulations. The CMIP5 models will become the main source of climate projections assessed by the International Panel on Climate Change (IPCC) in its 5th Assessment Report.
Two scenarios are used for the 56 ensemble members from 20 climate models in the CMIP5 archive: The representative concentration pathway (RCP) 4.5 future emission scenario, that stabilizes radiative forcing at 4.5 W m-2 in the year 2100, resulting in approximately 550 ppm of CO2 by 2100. The other scenario is the same CMIP3 20th century and future “business as usual” (SRES A1B emission scenario) model output. A1B attains CO2 levels of 750 ppm by 2100 and is hence a more aggressive scenario than RCP4.5. Models with more than 75% of their distribution falling outside the observed range of 6.13 to 8.43 million km2 were rejected. Of the 20 CMIP5 models, 17 models were retained, resulting in a total of 38 ensemble members.
Here's how things look for modeled, observed and projected September mean sea ice extent in the 1900-2100 period:
Figure 2. Time-series of modeled (colored lines) and observed (solid red line) September sea ice extent from 1900 to 2100. All 56 individual ensemble members from 20 CMIP5 models are included as dotted colored lines, with their individual model ensemble means in solid color lines. The multi-model ensemble mean is based on 38 ensemble members from 17 CMIP5 models (shown in black), with +/- 1 standard deviation shown as dotted black lines. Figure inset is based on the multi-model ensemble mean from CMIP5 and CMIP3, +/- 1 standard deviation.
I haven't drawn in this melting season's result, but it will obviously fall outside the 1 standard deviation of the CMIP5 multi-model ensemble. Some ensemble members are even lower, but that's because they started out much lower than the actual observations.
Results are described as follows in the paper:
Based on multi-model ensemble mean extents at the beginning of the 20th century, there is a 1.1 million km2 difference in the mean September extent between CMIP3 and CMIP5 (Figure inset, N.). During the period of satellite observations, the September CMIP5 multi-model ensemble mean tends to be slightly lower than the observed extent until 2007, after which it is higher. By contrast, the CMIP3 multi-model ensemble has a positive bias throughout the period of observations and especially during the most recent decade. Turning to the end of the 21st century, the CMIP5 multi-model ensemble mean never reaches ice-free conditions (defined here as less than 1.0 million km2), but the minus 1 standard deviation drops below the ice-free threshold around year 2045. Several CMIP5 models (CanESM2, GISS E2-R, GFDL-CM3, NCAR CESM, MIROC-ESM and ESM-CHEM) show essentially ice-free conditions by 2050, with the CanESM2 model having an ensemble member reaching nearly ice-free conditions as early as 2016 (0.54 million km2). By contrast, despite the more aggressive emission scenario (SRESA1B) driving the CMIP3 models, an overall more extensive sea ice cover is retained, with the minus 1 standard deviation reaching nearly ice-free conditions in 2075.
The same thing was done for March mean sea ice extent, with the following result:
The lower CMIP5 March extent compared to CMIP3 results in good overall agreement with the observations, though the observed values fall below the CMIP5 multi-model ensemble means in recent years.
Stroeve et al. also test whether the observed trend falls within the distribution of simulated trends at a specified level of statistical significance:
Turning to the modern satellite era, 1979-2011, more CMIP5 ensembles have a smaller rate of decline than observed, which at -0.84 million km2 per decade is nearly twice as large as the trend for 1953-2011. Forty-six of 56 ensemble members have trends outside of the 2 bound for the observations and 9 ensemble members have trends that are not statistically different from zero at the 90% confidence level. Although most model trends remain slower than observed, six ensemble members have rates of decline larger than observed. Overall, 64% of the ensemble member trends are statistically different from the observed trend at the 90% confidence level. In contrast, 85% of the CMIP3 ensemble members have trends that are statistically different from observed. The multi-model ensemble mean trend over the satellite period is -0.50 million km2 decade, which is 70% larger than the CMIP3 multi-model mean value of -0.35 million km2 decade.
It's clear that the CMIP5 models do a better job than the CMIP3 models used for the 2007 IPCC Fourth Assessment Report. But better is not the same as good. In this case 'less bad' would be a more accurate description. If Arctic sea ice continues melting at this rate, the question arises whether there is any sense in trying to optimize models. By the time they capture all the processes that are involved in melting, transporting and compacting the Arctic sea ice cover, it might already be gone. Maybe some of the money and human intelligence that go into models could be better spent on observational purposes (satellites, buoys, moorings) and the consequences that disappearing Arctic sea ice is already producing: changing weather patterns, methane release, permafrost melt and erosion.
As for the new IPCC report coming out next year: This time, a crazy melting season that shatters all previous records occurs just before the deadline and not after it. Will the seriousness and potential danger of an ice-free Arctic be reflected in the relevant chapters? Or will the IPCC again be too conservative (see this piece by Bryan Walker on the Hot Topic blog), hiding behind models that still haven't caught up with reality? The (lead) authors have their work cut out for them, but I'm sure they'll do a good job. We'll see what happens then.
Thanks goes out to Julienne Stroeve for allowing me to cover her important paper. Dr. Stroeve is currently in the middle of the ice pack to survey ice conditions and collect validation data (more on that later). To share her experiences she has started a blog: Ice Edge 2012