Efficient statistical forecasting using Deep-Learning Reduced Order Models

We propose to use data-driven Reduced Order Models (ROM) as a replacement to physicsbased subsurface flow models in statistical forecasting tasks. The utilized data-driven ROMs will be build using Deep-learning (DL) techniques. Deep-learning is a fast growing sub-field of machine learning where Deep Neural Networks (DNN) with several hidden layers (including convolutional and residual layers) are fitted to the training data. 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 build (i.e. learned) using a set of detailed 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 (i.e. simulation runs) close to the target distribution mimicking importance sampling techniques (aka. active learning). The proposed framework will build on recent advances in deep learning for real-time fluid simulation [1]. Unlike iterative solvers, DNNs have a fixed computational complexity, which will drive significant reduction in the computational cost of the extracted ROMs.

We plan to address the following task using the proposed framework: (a) real-time uncertainty propagation and inverse uncertainty quantification (i.e. history matching) of large scale subsurface reservoirs (b) tail-risk assessment of production forecast (c) robust production optimization while accounting for different sources of geological uncertainties. In order to address the significant computational needs to perform these tasks, we will rely on open source Deep Learning packages with active support for GPUs [2] to exploit the power of modern GPUs from high-level computing languages (i.e. Python). In some sense, we will build a hybrid CPU/GPU computing framework, where the ROM training data are generated using standard reservoir simulation packages running on CPUs while the ROMs are trained and evaluated on GPUs.