2020-10-28 ESMD/E3SM PI Meeting -- Ocean Breakout Session

Ocean Breakout (10/28/2020)

Topic: D3S2-BR#3
Time: Oct 28, 2020 03:15 PM Eastern Time (US and Canada)

Breakout Report:

Key missing or uncertain processes:

Mesoscale/submesoscale eddies
- Eddy Energy CPT examining in detail source and sinks of eddy energy from high resolution simulations and observations. Working to constrain an eddy energetics scheme and backscatter.
Non-Boussinesq and Nonhydrostatic, but questions remain when and where these capabilities are needed
SUMMARY: focus on scale aware parameterization is critical and determining what advances are needed in what regions.

Advances to leverage

ML equation discovery and physics informed NN can parameterize mesoscale eddy effects. Need to think about how to couple current codes with advanced ML/AI algorithms that are not FORTRAN.
- ML can expose issues with model DYCORE. For example it may expose the need for a more accurate advection scheme
Ensembles are needed to help evaluate models, can also use this capability to do some parameter optimization (simple ML) especially in coupled mode (assimilation in all components)
SUMMARY: Physics informed ML and ensembles hold promise for better understanding models and improving projections

How do we accelerate progress

Community involvement
- Even though E3SM is not a community model, we need input from outside researchers.
- Need to create an infrastructure that allows outside researchers to run the model without significant time and effort from E3SM staff
- Need clear documentation of the model and associated workflows
- For developments to come back to E3SM, clear standards of testing and design standards are necessary.
Levels of model complexity
- We can’t drive to high resolution alone
- Low resolution models allow for more rapid testing and development
  - Also can address different science questions (projections and uncertainty)
- Lower complexity models allow to isolate processes and test new ideas
- Work is needed to understand the minimum set of processes and feedbacks needed to understand the model.
- Need reduced set of BGC for testing and to move toward examination of BGC/ocean physics coupling
SUMMARY: investment is needed to make university involvement easier. An initial upfront cost but leads to lower cost over the longer term.

Notes:

Laure Zanna: Turbulence Closures in Ocean Models

Goals: improve parameterization of mesoscale eddies through energetics. Link momentum, buoyancy, eddy energy closures

Sources, sinks, and transfer of energy across scales is key. Mesoscales in particular are important; extract energy from mean flow and improves model output. GM mimics baroclinic instability and reduces spurious convection and mixing, but doesn’t account for eddy energy. Eddy energy can be used to inform GM coefficient, but then you’re still missing some energy pathways, so we need to rethink momentum closures. Bachman 2019 reinjects available potential energy from GM into resolved KE. Stocastic and non-newtonian closures are also an option. Overall, recent closures have focused on targeting energy transfers; have shown a reduction in biases in ocean transport. Challenge: how do we do this across resolution and in global models? Which momentum closure is best? What is the impact of vertical structure? There is a need for observationally-constrained and unified buoyancy and momentum closures, via energetics, for a robust scale and flow aware implementation in IPCC-class models

Questions:

Does tuning and validation require high resolution models since we don’t have energy cascade observations?

We are relying on bottom drag to account for flow around topography

Have non-newtonian closures been implemented? (a few in MOM6)

How to we figure out what the eddy length scale should be? How does that relate to GM kappa? (think about what is resolved and what is not resolved, but this is not clear cut)

Discussion:

What are the grand challenges in ocean/ice modeling? For example:
1. What are the key missing or more uncertain processes that should be prioritized for model development?
  1. Xylar: Boussinesq and non-hydrostatic. Need to think about both software engineering, but also think about the expense and where we need it.
2. What processes require more improvements in scale awareness?
  1. Mark: if we’re running at eddy permitting, what parameterizations should we be thinking about? Need to think about how we transfer from low to high res
  2. Laure: think about backscatter (stochastic methods, non-newtonian). Eddy permitting is still bad at energy transfer
  3. Anand: can run regional model if cheap and accurate
3. Is there a game-changing scale for modeling processes in each component?
4. Can we develop a heirarchy of models that actually informs coupled modeling efforts?
  1. Is there something between forced ocean sea ice and fully coupled runs?
    1. “pencil atmosphere”
    2. FAFMIP
    3. other?
  2. Zhengyu Lui: what is minimum set? He agreed this is a good idea, going from simple to fully coupled.
How can we overcome these challenges and accelerate progress?
1. Luke: better leverage community work, connections to cpt, use dsls to accelerate progress. Shouldn’t reengineer everything
2. Anand: have model that people can play with (e.g. single column mode)
3. Xylar: improvements in compass would help make it more accessable
4. Mark: post test cases after this release, improve readability of the code (deobfuscate!) but need to decide how much we can do
5. Zhengyu: Can we work with universities?
6. Wieslaw: community can contribute if there is good documentation, make process more user friendly
7. Andrew: there is a community/E3SM gap
8. Luke: this also makes things easier within E3SM
9. Zhengyu: how do we better communicate with the community?
10. Xylar: It’s time consuming for us to integrate others' work
11. Luke: we need very clear code standards so people can do what we expect
12. Kat: needs to be more communication for ocean BGC
13. Xylar: BGC analysis would improve visability
14. Luke: is there a reduced set of BGC that we can run?
What opportunities/recent advances can E3SM leverage?
How can we improve E3SM’s development and evaluation process? For example:
1. How could development better target known biases?
2. How can we ensure good coupled model behavior while developing component models?
3. How can development and evaluation be made more efficient?
machine learning:
1. Zanna: used equation discovery in machine learning. Equation discovery is easy, neural net is more difficult (but not impossible once you discover the parameters)
2. Anand: had a student who worked on machine learning. They found that you turn up things, like you need to have a more accurate advection scheme - you are loosing the curvature of the field, need to put that back in, as you change resolution.