5 ms·

I guess climate scientists don't use version control?

norswap 7y ago
The article isn't very precise, but my understanding was that the run some kind of machine learning based on past data. Adding in the more recent data changed the predictions dramatically. So it's not an issue of versioning (I assume they could reproduce the old results using the old data).
- whatshisface 7y ago
  It's not machine learning, it's a physical model. Computers solve a big differential equation that takes in all of the physical things we know about the Earth's climate system and runs them forward in time. The increase in thermal predictions are due to changes in the cloud and aerosol models which were made due to an improved understanding of the physics involved.
  - cultus 7y ago
    The line can be a bit blurry, because they are actually running inverse problems with the PDEs to fit the data. This needs regularization just like statistical ML models. Usually this amounts to solving an iterated least square problem, just like logistic regression. Bayesian methods are becoming more popular, though.
    - dls2016 7y ago
      > because they are actually running inverse problems with the PDEs to fit the data Could you explain this a bit more? Are you talking about data assimilation?
      - cultus 7y ago
        Yes, that's right. The issue is that systems like these have a lot of free parameters that have to be estimated from the data. This is basically because you have to make assumptions if you don't want to simulate Earth's systems down to the subatomic level. The inferred values of these parameters are usually very important as well for interpretation. All PDEs are solved numerically by discretizing them into linear or nonlinear systems. If it's convex enough, Tikhonov regularization is often used. Linear PDEs, where the solution depends only linearly on its derivatives, can be discretized into matrices. This gives a system like A m = d, where A is the discretized PDE, m is the "model" or parameter values, and d is the data. This is usually over or under-determined from the data, so it has to be solved by least squares. Since the resulting system is usually almost singular, regularization is applied, usually by penalizing the L2 norm of the model: min || A m -d || + r ||m||^2 If it's nonlinear but convex enough, one can just solve it by fixed point iteration, with the above linear equation being solved at every step (Newton's method often doesn't work well). If it's too nonconvex, then MCMC or other Bayesian methods are better, since a single solution is useless if the parameter values could plausibly around widely separated minima.
- peterwoerner 7y ago
  They technically use the term constitutive model development. The math is similar to machine learning. Typically we know the basics of the physics, i.e. conservation of energy, momentum and angular momentum which set up the equations. But there a number of constitutive laws which relate parameters between these are "machine learned" but in generally, they are typically linear relations.
gameswithgo 7y ago
it isn't that, it is just a badly written article.

The article isn't very precise, but my understanding was that the run some kind of machine learning based on past data. Adding in the more recent data changed the predictions dramatically. So it's not an issue of versioning (I assume they could reproduce the old results using the old data).

whatshisface 7y ago
It's not machine learning, it's a physical model. Computers solve a big differential equation that takes in all of the physical things we know about the Earth's climate system and runs them forward in time. The increase in thermal predictions are due to changes in the cloud and aerosol models which were made due to an improved understanding of the physics involved.
- cultus 7y ago
  The line can be a bit blurry, because they are actually running inverse problems with the PDEs to fit the data. This needs regularization just like statistical ML models. Usually this amounts to solving an iterated least square problem, just like logistic regression. Bayesian methods are becoming more popular, though.
  - dls2016 7y ago
    > because they are actually running inverse problems with the PDEs to fit the data Could you explain this a bit more? Are you talking about data assimilation?
    - cultus 7y ago
      Yes, that's right. The issue is that systems like these have a lot of free parameters that have to be estimated from the data. This is basically because you have to make assumptions if you don't want to simulate Earth's systems down to the subatomic level. The inferred values of these parameters are usually very important as well for interpretation. All PDEs are solved numerically by discretizing them into linear or nonlinear systems. If it's convex enough, Tikhonov regularization is often used. Linear PDEs, where the solution depends only linearly on its derivatives, can be discretized into matrices. This gives a system like A m = d, where A is the discretized PDE, m is the "model" or parameter values, and d is the data. This is usually over or under-determined from the data, so it has to be solved by least squares. Since the resulting system is usually almost singular, regularization is applied, usually by penalizing the L2 norm of the model: min || A m -d || + r ||m||^2 If it's nonlinear but convex enough, one can just solve it by fixed point iteration, with the above linear equation being solved at every step (Newton's method often doesn't work well). If it's too nonconvex, then MCMC or other Bayesian methods are better, since a single solution is useless if the parameter values could plausibly around widely separated minima.
peterwoerner 7y ago
They technically use the term constitutive model development. The math is similar to machine learning. Typically we know the basics of the physics, i.e. conservation of energy, momentum and angular momentum which set up the equations. But there a number of constitutive laws which relate parameters between these are "machine learned" but in generally, they are typically linear relations.

gameswithgo 7y ago

it isn't that, it is just a badly written article.