Data-driven Ocean Modelling

This thesis extends recent progress in data-driven weather forecasting by developing and assessing data-driven models of the ocean. Specifically we develop data-driven emulators of an idealised configuration of a physics-based ocean simulator, MITgcm. While work focuses on idealised cases, these simplified configurations nevertheless capture important aspects of the dynamics of the ocean, including coastal geometry. We develop data-driven emulators and analyse various aspects of these, namely their relationship to the ocean physics, their ability to predict near coast values, and their ability to iterate well. We assess the iterative performance of different data-driven models, and of different iteration approaches. First we consider explainable AI (X-AI) in the context of a simple regression-model, trained to emulate an idealised sector configuration of the MITgcm simulator. An important question around data-driven models is whether these models learn physically significant patterns, making them reliable to use in a variety of scenarios, or if they are learning non-physical statistical correlations, meaning they may perform badly outside of the precise statistical distribution within which they’ve been trained. We analyse the sensitivity of the regression model to its inputs through direct co-efficient analysis, and through ablation/withholding experiments. We see that the regression model predominantly performs in ways which match our domain expert knowledge. The model depends most heavily on the variables that we expect to be most important. When particular inputs are withheld, the model performance degrades most heavily in locations where we know these inputs dominate the dynamics. Work then looks at building a more complex data-driven model. Here we run a channel configuration of MITgcm and train a U-Net to emulate it. This one of the first applications of deep neural networks to emulate the ocean over short term prediction steps. We assess the emulator with a particular focus on the performance at coastal locations, as this is a specific challenge associated with the application of data-driven models to the ocean rather than the atmosphere. We see that the model performs very well overall, with errors low in comparison to the signal being forecast. However, when assessing performance only near land points, the model skill is comparable to a simple persistence forecast. This highlights a potential issue with data-driven forecasts — the optimisation used allows for systematic bias to occur, particularly if there are a limited number of challenging elements to the dynamics. We then assess the iterative performance of a U-Net ocean emulator. We show that when iterated over short term (2 week) forecasts, the model performs well, outperforming persistence and climatological forecasts. Errors are comparable to errors from an MITgcm ensemble, and fields look reasonable. Looking at longer forecasts, out to 3 months, we see that during the forecast period the model suffers from a sudden and total loss of skill. The predictions become physically implausible, and unbounded. Assessment of prediction-step sizes indicates this is somewhat related to step size, and hence to the number of iterations and to compounding errors, but that this alone is not a sufficient cause. A perturbation experiment shows that (prior to the breakdown in skill, and to some extent beyond it) errors in the U-Net compound and propagate at a similar level to that seen in the simulator. We see that particular elements of the physics bring about the loss of skill in many cases. We hypothesise that the breakdown in skill occurs when the predicted fields move outside of the scope of the training data. Finally we assess alternative data-driven models and iteration techniques to further understand the challenges associated with iterating data driven models. We see that while improving the long term stability of models is possible, it often comes at the expense of field fidelity, with a trade off between stability and ‘blurring’. Using more complicated methods, and improving the performance over a single time does not lead to improved iterative performance. A multi-network ensemble gives the best improvements for long term iterations, highlighting the promise of probabilistic data-driven methods. More generally, we note the optimisation-use dichotomy inherent in data-driven models. These models are inherently trained to meet the requirements of a simple loss function, however this is rarely in line with what they are used for. Forecast users are often concerned with whether a model is generalisable, and thus has learned the underlying physics of the situation. They also ideally want models which do not suffer from systematic bias. Lastly, to be truly useful models must provide high frequency output for a substantial forecast period, most easily achieved by using them iteratively, with predictions maintaining high spatial frequency, and realistic fields. However, very few of these (if any) are truly captured by the loss metrics used to train these models. We have seen outstanding performance and progress in the field of data-driven atmospheric models. Results from our idealised studies show promise for similar progress for data-driven ocean models. However, for both applications, we must note that the requirements of these models is not fully captured by what they are trained to do, and awareness of the limitations of these models is needed.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2024

Author dc:creator

• Furner, Rachel

Advisors dc:contributor.advisor

• Haynes, Peter
• Jones, Dani
• Munday, Dave
• Paige, Brooks
• Shuckburgh, Emily

• Data-driven
• Modelling
• Oceanography
• Ocean-modelling
• Machine-Learning

Licence

• https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/513c1658-a747-4f99-b494-5d96b7b9eef1/download
• https://creativecommons.org/licenses/by/4.0/

Language dc:language

eng

Author Identifier

0000-0003-0535-4388

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/377605