3 ms·

Imagine you are a climate scientist, and your mission is to save the world by working out how we are hurting it. You constantly run and re-try computer models

by jamez1 9y ago

Imagine you are a climate scientist, and your mission is to save the world by working out how we are hurting it. You constantly run and re-try computer models until you find one that you can publish. After 1000 different models you build one that seems to say there is some catastrophic effect caused by by humans with 99% confidence. The model then gets published, adds to the growing body of climate models that were made the exact same way your one was. Tell me - do you not see what is utterly incorrect with the above process?

godelski 9y ago
I do and and don't, but you are bringing up some good points. Unfortunately you're also demonstrating point two. An unfortunate limitation to my above statement is that it doesn't work so well with strangers on the internet. So instead let me talk about the scientific process, as a scientist myself. The above isn't how it works. But if it was, it might not even be wrong. Let's start with that. First off, why were they creating 1000 or even 1000's of models? You're on HN so I assume you program. In a big program how many times do you have to rewrite things before you get them working perfectly? Then how are these models accuracy determined? Well, with predictive power. The current weather models have high predictive power to 3 days (the European model has about 5). After that it is clear that error greatly increases. But that's weather, not climate, and I want to differentiate the two here. Climate modeling really is different from weather modeling. We have a lot of data on it, so it is more performing analysis to figure out the cause for changes. So let's ask ourselves how can we check the prediction power of our climate model? Well if our analysis is correct we can go about it two ways, or really check with both. We can use historic data (if you've ever looked into stock analysis, this is a common practice there). So we feed in our key variables and ask what temperature we get back. Does it match what we measured? And how well? We can also use it on new data, because as each day goes by we get more data. I'll stop there, as I don't want to write an entire paper. But a key part to science is its predictive power. If we don't have predictive power it isn't science, plane and simple. p.s. If you just plot the raw data (not modified for methods in which that data was gathered) you see a trend much worse than what current climate models predict. p.s.s Current models have held up in predictions over the past few decades. But there is always room for improvement.
mturmon 9y ago
You have, quite obviously, created a straw man. ("Tweak computer model until catastrophic outcome emerges.") This is not how influential climate science research works. Here's a paper (~200 citations) in JGR-A, by a colleague, to illustrate how it really works: http://onlinelibrary.wiley.com/doi/10.1029/2008JD010015/full http://onlinelibrary.wiley.com/doi/10.1029/2008JD010015/full Briefly, it's well-known that model representations of cloud ice are a significant source of uncertainty in climate modeling. This is because the ice scatters light upward out of the atmosphere, and can also trap emerging long-wave radiation from below. Cloud ice is hard to represent for many reasons - to mention just two, the particular form of the ice crystals (spherical or planar, termed "microphysical properties") matters, as does the vertical profile of the ice/water mixture in the cloud. How to improve their representation? One way is to measure cloud properties over years-to-decades time scales (which is now being done - no accident) and use this data to introduce the right cloud-process components into climate models. The paper surveys the measurements, proposes specific model improvements, and model-vs-data checks. As I said, I happen to work with the lead author of the paper above (although I'm not a climate scientist at all). He is, in fact, not twiddling knobs until catastrophic outcomes emerge. Sure, that's a logical possibility, i.e., someone could do that, but it does not correspond to reality by any stretch. Honestly, the notion is ridiculous, the gap between that notion and reality so large as to be hard to convey. Probably, at macro scale, the difference is that your straw man is purely open-loop (no checks on the knob-twiddling) whereas the one in the paper is brutally closed-loop (we will identify and make independent satellite measurements to quantify your model's accuracy). The paper describes working with others to design new measurements to get information about the key processes driving climate, so that the models can be improved and we can see what large-scale human forcing is doing to the climate system. Humanity is not a bit player any more, we're a driver, and we have to own up to that role.

by jamez1 9y ago

godelski 9y ago

I do and and don't, but you are bringing up some good points. Unfortunately you're also demonstrating point two. An unfortunate limitation to my above statement is that it doesn't work so well with strangers on the internet. So instead let me talk about the scientific process, as a scientist myself. The above isn't how it works. But if it was, it might not even be wrong. Let's start with that. First off, why were they creating 1000 or even 1000's of models? You're on HN so I assume you program. In a big program how many times do you have to rewrite things before you get them working perfectly? Then how are these models accuracy determined? Well, with predictive power. The current weather models have high predictive power to 3 days (the European model has about 5). After that it is clear that error greatly increases. But that's weather, not climate, and I want to differentiate the two here. Climate modeling really is different from weather modeling. We have a lot of data on it, so it is more performing analysis to figure out the cause for changes. So let's ask ourselves how can we check the prediction power of our climate model? Well if our analysis is correct we can go about it two ways, or really check with both. We can use historic data (if you've ever looked into stock analysis, this is a common practice there). So we feed in our key variables and ask what temperature we get back. Does it match what we measured? And how well? We can also use it on new data, because as each day goes by we get more data. I'll stop there, as I don't want to write an entire paper. But a key part to science is its predictive power. If we don't have predictive power it isn't science, plane and simple. p.s. If you just plot the raw data (not modified for methods in which that data was gathered) you see a trend much worse than what current climate models predict. p.s.s Current models have held up in predictions over the past few decades. But there is always room for improvement.

mturmon 9y ago

You have, quite obviously, created a straw man. ("Tweak computer model until catastrophic outcome emerges.") This is not how influential climate science research works. Here's a paper (~200 citations) in JGR-A, by a colleague, to illustrate how it really works: http://onlinelibrary.wiley.com/doi/10.1029/2008JD010015/full http://onlinelibrary.wiley.com/doi/10.1029/2008JD010015/full Briefly, it's well-known that model representations of cloud ice are a significant source of uncertainty in climate modeling. This is because the ice scatters light upward out of the atmosphere, and can also trap emerging long-wave radiation from below. Cloud ice is hard to represent for many reasons - to mention just two, the particular form of the ice crystals (spherical or planar, termed "microphysical properties") matters, as does the vertical profile of the ice/water mixture in the cloud. How to improve their representation? One way is to measure cloud properties over years-to-decades time scales (which is now being done - no accident) and use this data to introduce the right cloud-process components into climate models. The paper surveys the measurements, proposes specific model improvements, and model-vs-data checks. As I said, I happen to work with the lead author of the paper above (although I'm not a climate scientist at all). He is, in fact, not twiddling knobs until catastrophic outcomes emerge. Sure, that's a logical possibility, i.e., someone could do that, but it does not correspond to reality by any stretch. Honestly, the notion is ridiculous, the gap between that notion and reality so large as to be hard to convey. Probably, at macro scale, the difference is that your straw man is purely open-loop (no checks on the knob-twiddling) whereas the one in the paper is brutally closed-loop (we will identify and make independent satellite measurements to quantify your model's accuracy). The paper describes working with others to design new measurements to get information about the key processes driving climate, so that the models can be improved and we can see what large-scale human forcing is doing to the climate system. Humanity is not a bit player any more, we're a driver, and we have to own up to that role.