Sunday, 19 February 2012


Below is a very interesting comment by a senior Canadian climate scientist on how computer climate models should be tested.

"1. Why DO climate models have NO error bars (confidence intervals, say 95%) based on real observations?

2. I know the UK met office answer, that is: We have to just wait till say 2050 or 2100 to verify that "our models are right!"

3. NO that is NOT true! Let me explain;

The UK Met office and other climate modeling centres can do the following experiment:

Let them take their BEST AVAILABLE MODEL TODAY and let it be started in a PURE FORECAST MODE from about say 1950 or 1975.

yes it can be done! I did it for ocean wave models while at Env Canada , we call it a hindcast mode.

Let the UK met office run their best climate models from say 1950 till say 2010 (again in a purely fcst mode, NO adjustment of any kind) and then obtain fcst products say mean temp trends for five continents and three ocean basins and then do some error analysis using observed mean temp trend for five continents and three ocean basins for say 2010, 2011 and 2012.

Using these error stats, let them then obtain error bars for their fcst for 2050 and beyond.

My hunch is: the error bars will look like this:

For 2050: mean temp increase: 2C +/- 10C

For 2100; Mean temp incerase 3C +/- 15C ( something like these)

So you see these mean temp increase will be OF NO USE FOR ANY CLIMATE POLICY! Such fcsts will be TOTALLY USELESS! (right now all climate models give error bars based on so many climate model outputs, that is they compare model to model NOT model to real observations!)

Kindly allow me to give my personal experience:

Two years ago I gave a seminar on the GW science at York University here in Toronto. My seminar was arranged by my former good friend at Env Canada (Kaz Higuchi a good scientist and now on the Editorial Board of the Journal "Theoretical & Applied Climatology") and I gave a talk for over 40 minutes and during Q&A session, I said that "it would be useful for climate modelers to do this experiment and obtain error bars based on observed data"

Kaz told me that "the Canadian climate modelers DID EXACTLY THAT some few years ago and obtained such wide error bars they NEVER BOTHERED TO PUBLISH THOSE RESULTS!"

That is the problem with all climate models!

These modelers CAN do the experiment I suggest above BUT will NOT do it!"

I have just been told that in Chapter 8 of IPCC Report AR4, 2007 section "Metrics of Model Reliability" it says:
What does the accuracy of a climate model's simulation
of past or contemporary climate say about the accuracy of its projections of climate change? This question is just beginning to be addressed, exploiting the newly available ensembles of models."
Did you get that? "This question is just beginning to be addressed"! So now we know!

No comments: