Oceans play a key role in regulating climate.
Oceans play a key role in regulating climate. For instance, they absorb 93% of the extra heat generated by greenhouse gases whereas continents, ice and atmosphere all together only account for the remaining 7%. And that's why climatologists are so keen to work from the best possible representation of upper ocean dynamics when it comes to accurately predicting the extend of global warming for the next decades. But modeling with precision the energy exchange generated by ocean streams, tides, waves and wind flows remains a daunting challenge.
In essence, the problem is a matter of scale versus resolution. On the one hand, scientists have designed large-scale physical models that cover entire oceans over long periods of time but only at low-resolution. Given the vast geographical expanse and the long time span considered, computing these models at higher precision would require a whole life. On the other hand, oceanographers also build high-resolution models, but they only cover limited areas.
Compounding the issue is the fact that, although oceanographic models are based on causal physics and thus answer to a certain determinism, they also happen to be of a chaotic nature. Models are very sensitive to the original conditions. Tiny variations of these conditions can lead to quite different end results in short periods of time, thus bringing about some level of uncertainty.
A Long History of Collaboration
Reconciling the two types of models while also addressing the variability issue is one of the main goals of the ODYSSEY joint research team created in March 2022.
«It actually comes as the natural follow-up to two long-haul collaborations between our respective organizations. One between Ifremer and IMT Atlantique. The other one between Ifremer and Inria, sums up scientist Étienne Mémin, head of the new group at Inria. ODYSSEY will be a multidisciplinary group comprising 21 permanent members: numerical analysts, modeling specialists, data scientists, data assimilation specialists, AI researchers, so on and so forth. The overall goal is to come up with a new generation of oceanographic models that will better leverage the data. The current ones are built upon simplified hypotheses. Therefore, when it comes to prediction, they are not as efficient as we would like. We hope to bring them to the next level by leveraging two specific approaches. Namely: stochastic parametrization of models and machine learning.»
Addressing the Variability Issue
As far as exploiting stochastic methods is concerned, a milestone was reached in 2019 when Ifremer and Inria, together with Imperial College London, were granted funding by the European Research Council (ERC) in order to tackle the uncertainty issue in upper ocean fluid transport and introduce more variability in the models. This on-going 6-year European project is called STUOD.
The mathematical tools expected to come out of it will inject probability distributions of variables into the fine-grain patterns of local events and let them reverberate through out the large- scale model. Departing from a purely deterministic approach, the simulation will thus bring about a set of different realizations that describe the possible evolution of a phenomenon. Possible local fluctuations of water temperature, for instance.
Coupling Models and Data
The second approach regards the coupling of models and data. Hinging upon previous collaborations between Ifremer and IMT Atlantique, it comes against a backdrop where satellite observation, drifters and floats now provide growing amounts of information.
But we seldom access a complete picture. We are not in an ideal world where we would have high-resolution data covering the whole spectrum in terms of geographical expanse and time period. And if oceanographic models cover vast regions, they have low resolution. Therefore, a whole effort must be undertaken to connect these models to the available data. From this data, one can conduct a lot of physical analysis to get a more detailed understanding of certain processes such as the ocean-atmosphere system or the energy distribution mechanisms within the oceans. In this regard, our work will be conducted in conjunction with OCEANIX, a research chair in machine learning at IMT Atlantique.
Overall, Mémin concludes, «we are heading a toward a rather sophisticated combination of mathematical models, learning techniques and data assimilation methods for which researchers would be hard pressed to work in isolation. And that's why, from now on, I intend to be in Brest two days a week.»