Tuesday, February 01, 2005

Give Me That Old Time Science.

A distributed compute-engine simulation calculated a worst case temperature rise of 11C, making vast tracts of our planed uninhabitable (never mind that vast tracts of our planet are already uninhabitable). While I have absolutely no doubt that their simulations produced these results, I – unlike most of the population listening to this on the news – am fully aware of what simulations are and how they work.

It requires accurate initial conditions, accurate models, and an accurate physics engine.

So what does this mean?

As an example, a story from one of the grad students I met who was working on modeling a supernova explosion: we were chatting in the lab and I was fascinated by the process (they were running it on a Connection Machine -- a bunch of processors networked in a hypercube topology). He told me he had a physics model with over 200 parameters and the star had a fairly simple composition model, also with many parameters. He ran 4000 simulations before he got an explosion.

A climate model is an enormous undertaking.

1) You divide the planet into finite element. This is a long process, done offline where you must decide what portions of the atmosphere can be divided into 10 km voxels and what portions should be divided into 1km voxels. You then divide the surface into a grid and incorporate elevation data (which is important for prevailing winds). You divide the entire ocean into voxels.

2) You assign all the initial conditions, the density, composition, viscosity, fluid flow velocity, etc. Then you try to predict how it will change over time. There's a lot of guesswork here because you don't know what the exact composition of the ocean is or it's complete temperature profile, but you can get kind of close. Same problem with land cover – satellite data can give you percentage of plant cover, but you're really limited to visible light for absorption spectra, and you'll never know how land usage will change over time, so you project current trends. (We should all know how accurate curve fitted projections are outside the data set.)

3) You then project changes in the atmosphere due to human activity and apply adjustments throughout the run.

4) You develop physics models. Note that climate models are not the same as weather models. Weather models are HIGHLY non-linear and sensitive to initial conditions (chaotic) and can’t be used more than several days into the future before they diverge greatly. A climate model depends on long term averaging of weather chaotic behavior, taken primarily from history data. This too is difficult. We have perhaps 200 years of measured climate data, and aren't very sure about its accuracy or how it relates to worldwide surface temperatures (basically because nobody was measuring it back then.

Your physics engine becomes full of tweakable parameters. You tune it by attempting to make it behave as your (scant and full of assumptions) history data, based on initial conditions and population dynamics that are also largely guesswork.

5) You then run the simulations in Monte Carlo, varying your input parameters according to some algorithm that you also invent, and then run more simulations, varying even more parameters, in an attempt to find the most sensitive parameters, all the while hoping that you don’t get stuck in local minima. You focus on refining your most sensitive parameters and acquire more accurate data for initial conditions, while spending less time on parameters that have little effect on the outcome.

This critical process is Model Validation. You must be absolutely sure that your models accurately reflect what happens in real life to trust the simulation. For those of us who run simulations for a paycheck, the phrase is “your simulation is only as good as your model.”

From a critical eye, and being smart enough to actually understand it, I absolutely require seeing the data, the initial conditions, the parameters, and the process by which models are validated, because I know for a fact that you can make a simulation result in whatever the hell you want, using very subtle variations in sensitive parameters. I know. I've done it (in the past to get someone off my back about an issue I knew was a waste of time).

So these guys ran over 20,000 simulations (where they really need to run millions to truly validate the models and test for subtle combinations of parameters) and found the worst result out of these 20,000 simulations with different initial conditions and physics engine parameters results in the mean temperature of Earth rising 11C. You call a press conference. You get status boost at the university. You get people to give you grants to run bigger and better simulations. It's a big exciting time where everyone runs around talking with each other in heady conversation. It's WONDERFUL.

But, if you read the article, you will also see that their results varied much more widely than the results of previous investigator teams, meaning that while their worst case is really bad, their best case is really good (and this gets a lot less press because if there's no disaster, there's no excitement and no funding).

You get the attention of politicians. Poor countries LOVE global warming because that means that when the whole world starts rationing, you can sell your quota to the highest bidder (and it will be very high). So they all clamor over the Kyoto treaty and pitch it to whoever listens. Their profit motive moves them to talk the talk and run the walk and do whatever they can to assist these scientists that show the world going to hell (I really, really want to type “in a hand basket” – oops). Anyone with practical experience knows what happens to the boring results – they go into an obscure directory in the data set. The scary results end up on a Reuters feed, and everyone gets to have excited heady conversation and dynamics and stuff.

And none of this take’s into account that our technological capabilities are still growing at an exponential rate. It doesn't take into account that within mere decades, we'll be designing microbes and indeed complex organisms from scratch. Possibly we'll be designing some microbes that suck CO2 out of the air and process it into carbon nanotube fiber by the millions or even billions of tons (because that shit is a kick ASS construction material).

Truth is long before the Earth heats up enough to make the equator uninhabitable; we'll probably wipe out all life on Earth when some sociopath True Believer (TM) figures out how to design a MYDOOM virus that infects brown recluse spiders, causing them to seek out and bite evil Satanic American heretics. Then the evil Satanic American heretics in their last gasp release nanobots that seeks out True Believers (TM) and inserts a gene into their sperm stem cells that makes all of their kids pop out with blonde hair and blue eyes, thus causing them to murder all their family members in a holy rage.

So am I impressed that D. A. Stainforth, PhD got his star to explode? Not especially.

Give me an article about knockout gene mice any day and stem cells being used to cure diabetes. At least it has results that I can have confidence in.

Get back to me when these jokers show a clear path from me driving my Excursion across the street to get a pack of smokes to the extinction of the human race (ignoring for the moment that no species can stand the test of infinity (or in the case of us humans the test of the next few centuries because you can bet that whoever survives won’t resemble us much.))

Junk science pisses me off.

And I need a vacation.

No comments: