A hybrid four-dimensional variational data assimilation / scientific machine learning framework for coupled Arctic Ocean-sea ice model parameter calibration, state estimation, and nowcasting
Abstract
We propose to develop a computational and mathematical framework, within which fourdimensional variational data assimilation (4DVar) will be enhanced by scientific machine learning (SciML) methods to develop nowcasting capabilities of the coupled Arctic ocean-sea ice state. SciML is referred to here as the blending of data-driven machine learning with scientific computing. The approach will combine arguably one of the most advanced, adjoint-based data assimilation frameworks with new developments in the field of SciML (see the 2018 Department of Energy report on SciML). The conceptual framework rests on workflows that will harness several existing key components: (i) sparse, heterogeneous streams of in-situ observations of disparate types (ice tethered profilers, gliders, saildrones, various buoy programs, ship-based sections and fixed mooring arrays) that are currently being assimilated into regional Arctic ocean-sea ice analyses (nowcasts) and reanalyses; in addition, data from dedicated observational campaigns, such as ONRs Marginal Ice Zone (MIZ) and Stratified Ocean Dynamics in the Arctic (SODA) projects; (ii) satellite remote sensing observations, in particular of sea ice concentration and extent (from SSMI, AMSR), sea ice thickness retrievals (from SMOS, CryoSat-2, ICESat-2), sea ice motion and deformation (from RADARSAT and newly deployed SAR), ocean bottom pressure (from GRACE-FO), and sea surface salinity (from SMOS and Aquarius); (iii) a state-of-the-art coupled ocean-sea ice general circulation model (GCM) with several ice rheology solvers available; (iv) a 4DVar system that has been shown to be the most advanced system, capable of time-resolved assimilation of a wide range of satellite and in-situ observations, but that reaches limitations in the context of highly nonlinear sea ice parameterizations and ice rheology. We will demonstrate the capability of a hybrid 4DVar--SciML data assimilation scheme by addressing the most significant current shortcoming of the Arctic coupled ocean-sea ice nowcasting system: the sea ice dynamical solver. A SciML surrogate solver consisting of a neural network (NN) will be produced using training data from (i) synthetic data generated with the high-resolution HYCOM/CICE/NCODA modeling and assimilation system as well as regionally refined versions (in particular the US Navy Arctic Cap Nowcast/Forecast System; ACNFS), and (ii) high-resolution satellite SAR retrievals of sea ice motion and deformation. This component will be implemented into the regional Arctic 4DVar modeling system, and performance and uncertainty will be assessed via ensemble analysis, with open boundary conditions provided by the global HYCOM model. Next, with the goal of incorporating SciML into assimilation schemes, the response functions of the SciML dynamical solver NN will be deduced using backpropagation methods, and used along with the adjoint model to assimilate disparate date streams. Our work towards a 4DVar-SciML capability for nowcasting will pave the way forward for a more general coupling of SciML into predictive Earth system models and data assimilation schemes.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 31, 2020
- Source ID
- N000142012772
Entities
People
- Patrick Heimbach
Organizations
- Office of Naval Research
- United States Navy
- University of Texas at Austin