Title: Efficient statistical forecasting using Deep-Learning Reduced Order Models
We propose to develop data-driven Reduced Order Models (ROM) as a replacement to physics based subsurface flow models for statistical forecasting tasks. Data-driven ROMs will be build using Deep-learning techniques. Deep Neural Networks (DNNs) can eliminate the need for heavy feature engineer and are currently producing state-of-the-art results on several visual recognition tasks.
The proposed data-driven ROMs will be built using a set of full physics simulation runs representing the various sources of geological and model uncertainties. Further, the reduced order model will be iteratively enriched with more training samples close to the target distribution mimicking importance sampling techniques. The proposed framework will build on recent advances in deep learning for real-time fluid simulation and the recently introduced deep residual recurrent neural networks. Unlike iterative solvers, DNNs have a fixed computational complexity, which will drive significant reduction in the computational cost of the extracted ROMs.
Dr. Ahmed Elsheikh