Notes / Ideas / Progress

[andrewfagerheim]

This is an issue to keep track of meeting notes, progress updates, new ideas, etc. Hopefully this will help keep things organized as we work asynchronously. Other issues include:

• Project Assignment: requirements for the project as given by Prof Gentine
• References: readings, resources, repos

[andrewfagerheim]

Nov 10: Meeting with @dhruvbalwada @bernardw11

Notes:

• Large CNNs are expensive to evaluate, so it makes a course grid model less efficient (which is why they are preferred to high resolution models). Goal of this project is to 1) define a ML architecture that improves performance compared to physics parametrizations that is 2) roughly as computationally efficient as the physical parametrizations. The paper draft has created a CNN that fulfills the first point, maybe we could start thinking about the second.
• terms:
  • buoyancy is referring to the same thing as density
  • buoyancy frequency:
  • buoyancy flux:
• There are lots of forcings at the ocean surface which produce turbulence that would mix the entire ocean if this was the only type of force at play. However, the mixed layer depths are only ~100m deep, indicating there are restratification forces as well restoring density/buoyancy stratification.
• Restratification happens on the submesoscale (~10-20km), on scales that are sub-grid for large ocean/climate models. Therefore, the output is to calculate changes in buoyancy flux on a small grid and report this as a quantity for the model's course grid. The inputs are any large-scale parameters that influence buoyancy/density changes.

Possible efficiency improvements:

• Neural Net: NNs require less computational resources, so this could help on the efficiency side. Would it still produce a skillful set of predictions?
• Convolution Neural Network: Maybe it's possible to still use a CNN, but instead make it sparser to improve efficiency. Again, would this sacrifice skillful predictions?
• Inputs: Maybe it's possible to pare down the number of quantities that are input to the model, the paper started to address this with the ablation experiments (removing one input and assessing changes in R^2 values.

Next steps:

• [x] Dhruv will add dataset and current CNN to the LEAP hub
• [x] Bernard and I will review both and start to address the efficiency improvement opportunities

[andrewfagerheim]

Nov 29: Meeting with Abigail and @bernardw11

Notes:

• original regions is 15x15, then when processing it goes down to 10x10 degrees
• coarsening already weighs by area because the coordinate is degrees? I don't think I completely followed this discussion, Bernard might have more thoughts
• 2 measurements per day, averaged over 12 hour windows; this GCM isn't coupled (atmosphere is forcing at the beginning), 1/256 degree, hourly temporal resolution
• think of 40x40 as "the world," kernal size is 5x5 in first layer, then 3x3 in hidden layers