Earth, Atmosphere & Environment Sciences
Title : | MODES: Machine-learning for Ocean Data-assimilation, Estimation and Simulation |
Area of research : | Earth, Atmosphere & Environment Sciences |
Focus area : | Ocean science |
Principal Investigator : | Dr. Deepak Narayanan Subramanium, IISC Bangalore, Karnataka, Karnataka, Karnataka |
Timeline Start Year : | 2020 |
Timeline End Year : | 2025 |
Equipments : | (Laptop or
Desktop for the JRFs (2 no),
Assembled GPU Workstation
(with NVIDIA A6000 GPU)) |
Details
Executive Summary : | The coastal oceans and blue economy support livelihoods of tens of millions of people in India. By
various estimates, the coastal ocean contributes to 1% of the overall GDP of the country and is growing.
With such a high reliance on the coastal ecosystem, it is extremely important to monitor and predict the
health of our oceans with an emphasis on the impact of human activities. Ocean forecasts can be
utilized for informing sustainable decisions about human activities such as fishing, shipping, security and
surveillance that affect the coastal oceans. In this context, the MoES has invested in high quality ocean observation systems over the past several
years. Several ocean observation campaigns have also been undertaken giving us access to
unprecedented high-quality data. However, the processes of interest are multiscale and the data
available, even after all the best efforts, is but a sparse representation in space and time. As such,
estimating the initial and boundary conditions, parameters and closure models for numerical ocean
forecasting is laden with uncertainties. What is required now is to assimilate all the data collected over
the last several years and those being collected in real-time to improve high-resolution synoptic
forecasts of the ocean state. Moreover, a synergic blend of dynamical understanding with data is
needed using mathematically sound statistical machine learning principles. Additionally, the utility of
different observing systems must be established to finetune national observation plans. Towards this
end, data-driven dynamics-based feature models must be developed. The data assimilation methods
must respect the nonlinearity of the governing ocean primitive equations and the non-Gaussian, non-
stationary statistical distributions of the multiscale ocean features. To do so, the focus must shift from
estimating a covariance matrix from linearized or low-rank ensembles to new methods that accurately
describe the uncertainties in a dynamical subspace and propagate those using accurate numerical
schemes |
Total Budget (INR): | 54,46,720 |
Organizations involved