Efficient Forecasting of Drift Trajectories using Simplified Ocean Models and Nonlinear Data Assimilation on GPUs
MetadataShow full item record
This thesis presents research on efficient, massively parallel methods and algorithms related to short-term forecasting of drift trajectories in the ocean. The topic has clear societal applications and is an important tool for, e.g.,search-and-rescue operations at sea, planning of oil-spill cleanup, and collision detection between icebergs and offshore installations. In this work, we investigate computational techniques that can be used complementary to the operational methods already in place today. The traditional approach is to use complex ocean models, of which it is only feasible to run a small ensemble. Due to large uncertainties in initial conditions for oceanographic simulations, however, we propose to use simplified ocean models that capture the relevant physics on short time horizons. We base our simplified ocean models on the rotational shallow-water equations, simulated using an explicit, high-resolution, finite-volume scheme. Since such schemes can be implemented to run efficiently on the graphics processing unit (GPU), we can afford to run a large ensemble of simplified ocean models. Furthermore, we investigate nonlinear data-assimilation techniques, such as particle filters, that enable us to use available observations of the ocean state to reduce the uncertainty in the ensemble. Our hope is that this approach, possibly in combination with the operational methods, can give a more complete picture of the uncertainties in the forecasted drift trajectories. The thesis consists of an introductory part plus five scientific papers. The first two papers assess enabling technologies and methods needed for our approach to forecasting of drift trajectories. This includes evaluating numerical schemes for their suitability to capture oceanographic shallowwater flow, and assessing programming environments for GPU computing. The third paper presents a massively parallel algorithm for applying the recently proposed implicit equal-weights particle filter to a shallowwater model for forecasting of drift trajectories. In the fourth paper, we present a framework for running efficient oceanographic simulations using a modern finite-volume scheme initiated from operational ocean circulation forecasts. Finally, the fifth paper explores the possibility of using a very large ensemble with 10 000 members along with a basic particle filter method for ensemble prediction of drift trajectories.
Has partsPaper 1: Holm, Håvard Heitlo; Brodtkorb, André R.; Brostrøm, Gøran; Christensen, Kai Håkon; Sætra, Martin Lilleeng. Evaluation of selected finite-difference and finite-volume approaches to rotational shallow-water flow. Communications in Computational Physics 2020 https://doi/org/10.4208/cicp.OA-2019-0033 "First published in Communications in Computational Physics 2020, published by Global Science Press,"
Paper 2: Holm, Håvard Heitlo; Brodtkorb, André R.; Sætra, Martin Lilleeng. GPU Computing with Python: Performance, Energy Efficiency and Usability. Computation 2020 https://doi.org/10.3390/computation8010004 This is an open access article distributed under the Creative Commons Attribution License (CC BY 4.0)
Paper 3: Holm, Håvard Heitlo; Sætra, Martin Lilleeng; van Leeuwen, Peter Jan. Massively parallel implicit equal-weights particle filter for ocean drift trajectory forecasting. Journal of Computational Physics: X 2020 ;Volum 6 https://doi.org/10.1016/j.jcpx.2020.100053 Under a Creative Commons license (CC BY 4.0)
Paper 4: Brodtkorb, André Rigland.;Holm, Håvard Heitlo Real-World Oceanographic Simulations on the GPU using a Two-Dimensional Finite Volume Scheme Preprint: arXiv:1912.02457, 2019
Paper 5: Holm, Håvard Heitlo; Sætra, Martin Lilleeng; Brodtkorb, André Rigland. Data Assimilation for Ocean Drift Trajectories Using Massive Ensembles and GPUs