Water masses have historically been defined using certain potential temperature and salinity bounds for each given one. In view of projected changes in heat content and freshwater fluxes, will the bounds defined in the mid-20th century still hold?
Other approaches, based on geometric consideration (see Juza et al. https://www.sciencedirect.com/science/article/pii/S0924796318302173), have been proposed. The goal here is to investigate if significant water mass properties changes are found between late 20th versus late 21th centuries) using T/S diagrams, as well as changes in water masses total volume. And if a significant change in temperature is detected, estimate the heat content change in some water masses of interest. There is room for discussion about what criterion can/should be used for water mass description.
Participants can focus on one (or more) water mass of particular interest to him/her.
Building T/S diagrams for multiple CMIP ocean models and compare 20th and 21th centuries, detect potential changes.
Compute water mass volume evolution (traditional criterion and maybe new one(s))
Estimate changes in heat content.
Anticipated Data Needs
CMIP (5/6) potential temperature and salinity
Anticipated Software Tools
pangeo software stack (xarray/dask)
computation of water mass volume and heat content is a 3 for loop, 4 if statements that can be optimized for performance with f2py or numba.
Beginner/Intermediate python users are welcome, basic knowledge of observational oceanography (can read T/S diagram).
On the science side, I have been thinking about this problem quite a bit. How do you properly compare to watermass distributions? What is the proper distance metric.
I believe this can be answered through optimal transport theory:
The idea is to figure out how much work it takes to transform one distribution into another. This problem can be solved efficiently, and there is a cool python package for doing it
Particular to the ocean T/S problem is the question of the cost of moving water in T/S space. One could imagine an energy-based metric, which penalizes diapycnal transformations in favor of isopycnal ones.
@rabernat xhistogram seems like the kind of tool that will be very useful for the projects. I think it’s also a good idea to merge with Spencer’s project.
I guess the question here is how may people are gonna be working on this hack
and can we try to distribute the efforts in a way where everybody got a bit to chew on while keeping the idea of a working group.
I’m not super familiar with the format so we’ll need the organizers to help out with this. But this looks like some great fun coming up, looking forward to it!
I am interested in this project. We have been looking at LSW and finding it rather difficult to pin point the density ranges that define this water mass. I suspect this is the case for other water masses as well. Certainly a group exercise. Thanks for posting this.
@aromanou - I was reading a paper by JB Sallee on water masses of the Southern Ocean in CMIP5 historical runs (Sallee et al 2013). It said that they had to manually identify the density ranges corresponding to the water masses for each individual model, as they were not necessarily the same due to density biases between models and obs.
Is anyone still working on this? I’m keen to jump in to come with something up. There is few works for classifying water masses from argo profiles by Maze et al (2017) (using ML unsupervised classification, that I feel it will have to be adapted for cmip) and also other options like geometry-based detection approach by Juza et al (2018) that might give us a hint on where to start…
Hey Denise, I don’t think anyone is pursuing this further. It would be great if some unsupervised technique could be implemented to identify watermasses between different simulations, and maybe also if there was a way to do some sort of automatic inter comparison. So, that questions like - while the two models don’t have the same T/S properties for say intermediate waters, but can we still compare the heat/carbon uptake in the respective intermediate waters.
I’m working on this right now. The first approach will probably follow some of the PCM-adapted principles (Gaussian Mixture Model) from Maze et al. (2017) but it will be interesting to try a few other approaches to see how they respond to different watermasses unsupervised classification methods. Although I am not sure I’ll have time for that though. As soon as I get something working clean I will be happy to share it here.
I think the intercomparison would be another separated project (which I am also interested in) as there are already watermasses metrics (from Beadling, Russel, Heuzé…) and score skills (from Glecker, Parding… sorry I don’t recall the exact references here) with good agreement we could follow for climate models.