University of Toronto (and colleagues) sub-team: internal tides+(sub)mesoscale interactions through SWOT’s lens
Sorry for making you wait and I hope it will be worth it. And thank you Ryan for making this happen. And sorry for the lack of URLs in my references. I cannot have more than two, and had to remove a bunch.
Project Goals and Approach
Our project is mostly funded by the Canadian Space Agency [14SUSWOTTO] and the Natural Sciences and Engineering Research Council of Canada [RGPIN-2015-03684]. Because several grants overlap, our goals have somewhat grown, but the main focus is to address the following question: how can we isolate the internal tide signature present in a given SSH snapshot? This is likely to be important for SWOT data, whose resolution will go beyond mesoscale and mode-2 internal tide resolutions. Consequently, it is likely to access regimes in which the increasing non-stationarity of the scattering (sub)mesoscale flow will make standard internal tide retrieval procedures hard or impossible to apply. And even if my hypotheses eventually prove to be wrong, it will remain a fun question. We are starting to provide (or aiming to; it is a big baby!) partial answers on two fronts (pun intended).
- During her post-doc, Han Wang will test a deep-learning approach we have been toying with on-and-off for ~1.5 years to extract the internal tide signal from synthetic SSH snapshots. To do so, we have been leveraging a small numerical dataset with sufficient temporal resolution to extract the tidal signal with frequency filtering, and train the algorithm with both the raw and tidally-filtered SSH snapshots. Although our method is completely different, we are essentially after a product that would look similar to recent work by Torres et al. (2019).
- The PhD thesis of Jeffrey Uncu takes a diametrically different approach from the first one: instead of trying to diagnose the product of an interaction we do not understand well, we try to understand very well the process without worrying about applicability to SWOT products in the next few years. In particular, Jeff is studying how internal tides and (sub)mesoscale motions interact. His goal is to have something fundamental to say about the submesoscale range, which as of 2020 appears inaccessible to standard analytical methods (asymptotics, etc). He is mostly using single-layer shallow-water simulations. We expect these interactions to be relevant for the coastal regime, which is rich in submesoscale motions and high-mode internal tides.
- Han graduated from NYU last Summer, and was working on the wave-vortex decomposition of Bühler, Callies & Ferrari (2014). In particular, she devised a way to apply it to Lagrangian observations (cue to @selipot), which is in revision, and we have been wondering about ways to make it useful for SWOT on the side. She joined our group in October 2020.
- Jeff is a Ph.D. candidate in my group, and has started to work on it in 2018.
- Alice Nuz and Michael Poon were two undergraduate students who actually came up and elaborated on the DL method, Han has been using so far.
- My role is to try to keep up and be cheerful. I also prepare a half-decent coffee.
We are also collaborating with Hesam Salehipour (soon to be at WHOI) for the technical DL aspects, @apatlpo provided the data we have been working with in Han’s project and with whom we have been discussing theoretical aspects of Jeff’s project since the beginning. Jody Klymak (University of Victoria in BC) is co-PI on the CSA grant, with an observational aspect somewhat outside the scope of this post.
We are still working on the DL algorithm, which we will make public on GitHub along with whatever publication we submit. Our timeline for submission is in a few months, and because it took us a while to zero in on an architecture, I ask for your patience regarding the specifics.
We have been testing our DL algorithm on re-entrant channel simulations with ROMS, created and provided by Aurélien, and curated by myself (it was the object of the original tweet, available here). This output contains surface expressions of relevant quantities (e.g., SSH, SST, surface vertical vorticity…) and mixed mode-1 internal tide signals with a QG-ish turbulent flow growing out of a baroclinicially unstable jet. Each field comes in three flavours: raw snapshot, internal tide (derived by fitting the signal to tidal-frequency sin/cos functions, and a low-passed equivalent (see Ponte & Klein 2015).
In the future, and provided the method passes enough tests, we intend to apply it to more comprehensive and realistic datasets, such as those that folks in this thread are using (MITgcm, etc.). Some sort of fast-slow decomposition would need to be either available, or doable (i.e., high enough time sampling, but from the looks of it, everyone here seems to be working with such data). We do have access to a nested MITgcm simulation output that Brian Arbic and Dimitris Menemenlis ran in Toronto (I was part of one study that came from this collaboration, see Nelson et al. 2020), but would rather use datasets that are widely used, or have the potential to be.
Jeff is mostly using his brain, single-layer shallow water simulations with Dedalus and scaling arguments.
Possible Collaboration Points
Our DL method will not be directly applicable to SWOT unless we significantly modify it, train it with very realistic data, and uglify it with the SWOT simulator. Each of these steps would be a significant endeavour, and we could use outside help.
As for the project of Jeff, theoretical work is hard to plan, but if people have been thinking about this problem, we’d be happy to share our thoughts. We have found that the dynamics in this regime are extraordinarily rich, and there is no way we could write the same paper twice on this.
And about the data challenges: sure, why not. My only concern is that Han and Jeff are pumped, so, I just want to make sure that everyone is OK with the prospect of losing.