Back to search

IKTPLUSS-IKT og digital innovasjon

GPU Ocean - heterogeneous computing of drift in the ocean

Alternative title: GPU Hav

Awarded: NOK 9.0 mill.

It is important to predict the drift trajectories of oil spills, ice bergs, and other floating objects to protect the marine environment and for safe offshore operations. Modern numerical circulation models are sophisticated, and have good representations of the physical processes that drive the ocean circulation. The ocean currents are highly variable on short temporal and spatial scales, however, and small deviations in the models quickly develop into large errors. The models can be corrected using observations, but unfortunately there is a lack of direct observations of the ocean circulation, hence the predictions are often associated with large uncertainties. Model ensembles, that is, many simultaneous model simulations with slightly different forcing and initial conditions, can be used to quantify the uncertainties. Large spread between the different simulations indicate large uncertainty and vice versa. Today's numerical models are computationally demanding, however, and in practice the number of simulations in the ensemble is often too small. In this project we will make ensembles with up to thousands of simultaneous simulations, using simplified ocean circulation models and advanced supercomputing techniques. Such ensembles enable us to make more robust uncertainty estimates, and also provide information about the physical processes that dominate the uncertainties. Very large ensembles can also greatly benefit from observations. We can pick and choose those simulations that are dynamically consistent with the few observations that are available, hence we obtain more accurate predictions of the drift. To facilitate rapid prototyping, a GPU-accelerated Python code base has been developed. We have developed a complete and functional model and data assimilation system. The implemented data assimilation methods are state of the art, which will allow us to compare different combinations of data assimilation methods and model ensemble sizes. Furthermore, we have established a set of prototypes and examples demonstrating the use of the project code base. These have been used to conduct experiments on several different platforms, from laptops to supercomputers. Every publication in the project is accompanied by the relevant code base and data. The last version of the code base is always available as free and open source software on Github. The project has an ongoing cooperation with researchers at the University of Reading, which are international leading experts in data assimilation.

The project has contributed to the knowledge and competency on simplified ocean models, data assimilation and drift trajectory models on massively data-parallel hardware, for both the project partners and the international collaborator. This has spurred a wider activity on GPU computing at the Norwegian Meteorological Institute, including a two-day short course on GPU computing with participants from all parts of the institute and procurement of GPU-accelerated nodes for the in-house HPC system. Furthermore, there are applications within the operational ocean forecasting in which the results from the project is ready to be applied, e.g., storm surge forecasting. The highly interdisciplinary nature of the project has brought together experts in many different scientific fields, and formed lasting collaborations and networks that has already resulted in a successful RCN proposal.

Numerical ocean models are used for oil spill tracking, search and rescue, and in costly offshore operations involving large floating structures. The natural variability in the ocean currents is large, and a major problem is the lack of ocean observations that can be used to constrain the ocean model. Ongoing efforts focus on exploiting advanced data assimilation techniques and devising efficient observation sampling strategies, but traditional ocean modeling systems are computationally demanding and the full range of forecasting uncertainties is rarely explored. We propose to complement the traditional ocean modeling systems with a massive ensemble of simplified ocean and drift models. These simplified models will be combined with fully non-linear data assimilation and the codes will be executed on massively parallel architectures. We aim to use simplified physical and mathematical models that are valid for the short predictive time scales necessary for operational oceanography, which is typically 1-3 days. These simplified models will be initialized from the traditional modeling system and hence provide incremental updates to the full description of the physical state of the ocean. The result is accurate ocean current predictions with detailed uncertainties.

Publications from Cristin

No publications found

Funding scheme:

IKTPLUSS-IKT og digital innovasjon