Confirmation of Richardson's power laws to make models better

Jody Radzik · 11-04-2009, 02:58 PM

http://www.newscientist.com/article/...ls.html?page=1

These findings might solve the problem of parameterisations, making the models much more granularly accurate.

Mike Kovalchick · 11-05-2009, 07:00 AM

EDIT: Remove. Existing new thread here:
http://www.stormtrack.org/forum/showthread.php?t=22241

Interesting article that suggests an excellent possibility that forecast models can be vastly improved using fractals.

http://www.newscientist.com/article/...html?full=true

..."The results point to a new view of the atmosphere as a vast collection of cascade-like processes, with large structures the size of continents breaking down to feed ever-smaller ones, right down to zephyrs of air no bigger than a fly...."

..."The implications promise to transform the way we predict everything from tomorrow's local weather to the changing climate of the entire planet. "We may never be able to view the atmosphere and climate in the same way again," says team member Shaun Lovejoy of McGill University in Montreal, Canada...."

David Wolfson · 11-05-2009, 11:05 AM

Boiled down, the essential point is that many random (in the sense of stochastic) processes are self-dependent over a wide range of scales rather than composed of independent elements following a Random probability distribution. Many statistics of such processes are defined by a skewed "fat tail" distribution rather than the Gaussian bell curve.

The multifractal components of the Earth's atmospheric circulation are clear, i.e. "understood", when you look at the distribution of the model ensemble runs for example. The result of stochastic initialization is a lot more variation and skew than a Random distribution. However the ensemble mean is still an attractor for practical purposes. Theoretically it's possible for a butterfly in the west Pacific to result in an ice-ball Earth but the probability is vanishingly small.

The real question I see that is posed to the modelers is whether some initialization parameters to Bayesian/Monte-Carlo ensembles are incorrectly generated. I'm not in the know on this, but I suspect that this is being considered, because very smart people are involved. For example, snow/ice cover isn't distributed randomly, so its parametric values in Monte-Carlo simulation shouldn't be generated from a Random distribution. FWIW.

11-04-2009, 02:58 PM	#1
Jody Radzik Member Join Date: Mar 2007 Location: Northern California Posts: 159	Confirmation of Richardson's power laws to make models better http://www.newscientist.com/article/...ls.html?page=1 These findings might solve the problem of parameterisations, making the models much more granularly accurate. __________________ Jody Radzik WX1NMU mesophile.org

11-05-2009, 07:00 AM	#2
Mike Kovalchick Member Join Date: Jan 2005 Location: Lansing/Gaylord, Michigan Posts: 170	Tomorrow's weather: Cloudy, with a chance of fractals EDIT: Remove. Existing new thread here: http://www.stormtrack.org/forum/showthread.php?t=22241 Interesting article that suggests an excellent possibility that forecast models can be vastly improved using fractals. http://www.newscientist.com/article/...html?full=true ..."The results point to a new view of the atmosphere as a vast collection of cascade-like processes, with large structures the size of continents breaking down to feed ever-smaller ones, right down to zephyrs of air no bigger than a fly...." ..."The implications promise to transform the way we predict everything from tomorrow's local weather to the changing climate of the entire planet. "We may never be able to view the atmosphere and climate in the same way again," says team member Shaun Lovejoy of McGill University in Montreal, Canada...." Last edited by Mike Kovalchick; 11-05-2009 at 07:08 AM.

Thread Tools
Show Printable Version Email this Page
Display Modes
Linear Mode Switch to Hybrid Mode Switch to Threaded Mode

11-05-2009, 11:05 AM	#3
David Wolfson Member Join Date: Mar 2004 Location: Phoenix, AZ Posts: 1,779	Boiled down, the essential point is that many random (in the sense of stochastic) processes are self-dependent over a wide range of scales rather than composed of independent elements following a Random probability distribution. Many statistics of such processes are defined by a skewed "fat tail" distribution rather than the Gaussian bell curve. The multifractal components of the Earth's atmospheric circulation are clear, i.e. "understood", when you look at the distribution of the model ensemble runs for example. The result of stochastic initialization is a lot more variation and skew than a Random distribution. However the ensemble mean is still an attractor for practical purposes. Theoretically it's possible for a butterfly in the west Pacific to result in an ice-ball Earth but the probability is vanishingly small. The real question I see that is posed to the modelers is whether some initialization parameters to Bayesian/Monte-Carlo ensembles are incorrectly generated. I'm not in the know on this, but I suspect that this is being considered, because very smart people are involved. For example, snow/ice cover isn't distributed randomly, so its parametric values in Monte-Carlo simulation shouldn't be generated from a Random distribution. FWIW.